CONTENTS PREFACE TO"MEANS AND THEIR INEQUALITIES" .............. . Xlll PREFACE TO THE HANDBOOK ..................................... . XV .. BASIC REFERENCES ............................................... . XVll . NOTATIONS .......................................................... . XlX . 1. Referencing ....................................................... . XlX . 2. Bibliographic References ........................................ . XlX . 3. Symbols for Some Important Inequalities ....................... . XlX 4. Numbers, Sets and Set Functions ................................ . XX 5. Intervals . . . . . . . . . . . . . . . . . . . . . . . . . . .............................. . XX . 6. n-tuples ........................................................ . XXl .. 7. Matrices ....................................................... . XXll .. 8. Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . XXll 9. Various . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xx111 A LIST OF SYMBOLS .. .. .. . . . . . .. . .. . . . .. .. . . .. . . . . .. . .. .. . . . .. . .. . . xxiv AN INTRODUCTORY SURVEY . .. .. . .. .. . .. .. .. .. .. .. .. .. . .. . .. .. .. . xxvi CHAPTER I INTRODUCTION . .. .. .. . .. . .. .. .. . .. .. . .. . . .. .. . .. . .. . 1 1. Properties of Polynomials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 1.1 Some Basic Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 1.2 Some Special Polynomials ....................................... 3 2. Elementary Inequalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4 2.1 Bernoulli's Inequality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4 2.2 Inequalities Involving Some Elementary Functions . . . . . . . . . . . . . . . . 6 3. Properties of Sequences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 3.1 Convexity and Bounded Variation of Sequences . . . . . . . . . . . . . . . . . . 11 3.2 Log-convexity of Sequences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16 3.3 An Order Relation for Sequences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21 4. Convex Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25 4.1 Convex Functions of Single Variable . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25 4.2 Jensen's Inequality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30 4.3 The Jensen-Steffensen Inequality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37 4.4 Reverse and Converse Jensen Inequalities . . . . . . . . . . . . . . . . . . . . . . . . 43 4.5 Other Forms of Convexity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48 4.5.1 Mid-point Convexity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48 4.5.2 Log- convexity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48 4.5.3 A Function Convex with respect to Another Function . . . . . . . . 49
..
Vll
4. 6 Convex Functions of Several Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . 50 4. 7 Higher Order Convexity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54 4.8 Schur Convexity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57 4. 9 Matrix Convexity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58 THE ARITHMETIC, GEOMETRIC CHAPTER II AND HARMONIC MEANS . . . . . . . . . 60 1. Definitions and Simple Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60 1.1 The Arithmetic Mean . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60 1.2 The Geometric and Harmonic Means . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64 1.3 Some Interpretations and Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . 66 1.3.1 A Geometric Interpretation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66 1.3.2 Arithmetic and Harmonic Means in Terms of Errors . . . . . . . . . 67 1.3.3 Averages in Statistics and Probability . . . . . . . . . . . . . . . . . . . . . . . 67 . . . . . . . . . . . . . . . . . . . . . . . . . . 68 1.3.4 Averages in Statics and Dynamics 1.3.5 Extracting Square Roots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68 1. 3. 6 Cesaro Means . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69 1. 3. 7 Means in Fair Voting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70 1.3.8 Method of Least Squares . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70 1.3.9 The Zeros of a Complex Polynomial . . . . . . . . . . . . . . . . . . . . . . . . . 70 2. The Geometric Mean-Arithmetic Mean Inequality . . . . . . . . . . . . . . . . . . . 71 2.1 Statement of the Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71 2. 2 Some Preliminary Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72 2.2.1 (GA) with n == 2 and Equal Weights . . . . . . . . . . . . . . . . . . . . . . . . . 72 2.2.2 (GA) with n == 2, the General Case . . . . . . . . . . . . . . . . . . . . . . . . . . 76 2.2.3 The Equal Weight Case Suffices . .. .. . .. .. . .. .. . .. .. . .. .. .. .. 80 2.2.4 Cauchy's Backward Induction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82 2.3 Some Geometrical Interpretations 2.4 Proofs of the Geometric Mean-Arithmetic Mean Inequality . . . . . . . 84 . . . . . . . . . . . . . 85 2.4.1 Proofs Published Prior to 1901. Proofs (i)-(vii) 2.4.2 Proofs Published Between 1901 and 1934. Proofs (viii)-(xvi) . . 89 2.4.3 Proofs Published Between 1935 and 1965. Proofs (xvii)-(xxxi) ... 92 2.4.4 Proofs Published Between 1966 and 1970. . . . . . . 101 Proofs (xxxii)-(xxxvii) 2.4.5 Proofs Published Between 1971 and 1988. Proofs (xxxviii)-(lxii) 104 . . . . . . . . . 114 2.4:6 Proofs Published After 1988. Proofs (lxiii)-(lxxiv) 118 2.4.7 Proofs Published In Journals Not Available to the Author 119 2.5 Applications of the Geometric Mean-Arithmetic Mean Inequality 2.5.1 Calculus Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119 2.5.2 Population Mathematics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120 2.5.3 Proving other Inequalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120 2.5.4 Probabilistic Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124 125 3. Refinements of the Geometric Mean-Arithmetic Mean Inequality 3.1 The Inequalities of Rado and Popoviciu . . . . . . . . . . . . . . . . . . . . . . . . 125 3.2 Extensions of the Inequalities of Rado and Popoviciu . . . . . . . . . . . 129 3.2.1 Means with Different Weights . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132 3.2.2 Index Set Extensions Vlll
3.3 3.4 3.5 3.6 3.7
A Limit Theorem of Everitt .................................. . N anjundiah's Inequalities ..................................... . ................................. . Kober-Diananda Inequalities Redheffer's Recurrent Inequalities ............................ . The Geometric Mean-Arithmetic Mean Inequality with General Weights ....... . 3.8 Other Refinements of Geometric Mean-Arithmetic Mean Inequality 4. Converse Inequalities ............................................. . 4.1 Bounds for the Differences of the Means ....................... . 4.2 Bounds for the Ratios of the Means ........................... . 5. Some Miscellaneous Results ...................................... . 5.1 An Inductive Definition of the Arithmetic Mean ............... . 5.2 An Invariance Property ....................................... . 5.3 Cebisev's Inequality ........................................... . 5. 4 A Result of Diananda ......................................... . 5.5 Intercalated Means 5.6 Zeros of a Polynomial and Its Derivative ....................... . 5. 7 Nanson's Inequality .......................................... . 5.8 The Pseudo Arithmetic Means and Pseudo Geometric Means 5.9 An Inequality Due to Mercer .................................. . CHAPTER III THE POWER MEANS ............................ . 1. Definitions and Simple Properties ................................ . 2. Surns of Powers .................................................. . 2.1 Holder's Inequality ............................................ . 2.2 Cauchy's Inequality ........................................... . 2. 3 Power sums . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2. 4 Minkowski' s Inequality ........................................ . 2.5 Refinements of the Holder, Cauchy and Minkowski Inequalities 2.5.1 A Rado type Refinement .................................. . 2. 5. 2 Index Set Extensions ...................................... . 2.5.3 An Extension of Kober-Diananda Type .................... . 2.5.4 A Continuum of Extensions ............................... . 2.5.5 Beckenbach's Inequalities .................................. . 2. 5. 6 Ostrowski's Inequality ..................................... . 2.5.7 Aczel-Lorentz Inequalities ................................. . 2.5.8 Various Results ........................................... . 3. Inequalities Between the Power Means ............................ . 3.1 The Power Mean Inequality ................................... . 3 .1.1 The Basic Result ....................................... . 3.1.2 Holder's Inequality Again ................................. . 3.1.3 Minkowski's Inequality Again ............................. . ..... 3.1.4 Cebisev's Inequality ...................................... . 3.2 Refinements of the Power Mean Inequality ..................... . 3.2.1 The Power Mean Inequality with General Weights ......... . 3.2.2 Different Weight Extension ................................ . 3.2.3 Extensions of the Rado-Popoviciu Type ................... .
.
IX
133 136 141 145 148 149 154 154 157 160 160 160 161 165 166 170 170 171 174 175 175 178 178 183 185 189
192 192 193
194 194
196 198 198
199 202 202 203
211 213 215 216 216 216 217
3. 2. 4 Index Set Extensions ...................................... . 220 3.2.5 The Limit Theorem of Everitt ............................. . 225 3.2.6 Nanjundiah's Inequalities ................................. . 225 4. Converse Inequalities ............................................ . 229 4.1 Ratios of Power Means ........................................ . 230 4.2 Differences of Power Means ................................... . 238 4.3 Converses of the Cauchy, Holder and Minkowski Inequalities 240 5. Other Means Defined Using Powers ............................... . 245 5.1 Counter-Harmonic Means ..................................... . 245 5.2 Generalizations of the Counter-Harmonic Means ............... . 248 5.2.1 Gini Means ............................................... . 248 5.2.2 Bonferroni Means . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251 5.2.3 Generalized Power Means ................................ . 251 5.3 Mixed Means ................................................. . 253 ............................................. . 256 6. Some Other Results 6.1 Means on the Move . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 256 6.2 HlaV\rka-type inequalities ......................................... . 258 6.3 p- Mean Convexity ............................................. . 260 6.4 Various Results ............................................... . 260 CHAPTER IV QUASI-ARITHMETIC MEANS .................... . 266 1. Definitions and Basic Properties .................................. . 266 1.1 The Definition and Examples ................................. . 266 ........................... . 271 1.2 Equivalent Quasi-arithmetic Means 2. Comparable Means and Functions ................................ . 273 3. Results of Rado-Popoviciu Type .................................. . 280 3.1 Some General Inequalities ..................................... . 280 3.2 Some Applications of the General Inequalities ................. . 282 4 Further Inequalities ............................................... . 285 4.1. Cakalov's Inequality .......................................... . 286 4.2 A Theorem of Godunova ...................................... . 288 4.3 A Problem of Oppenheim ..................................... . 290 4.4 Ky Fan's Inequality ........................................... . 294 4.5 Means on the Move ........................................... . 298 5. Generalizations of the Holder and Minkowski Inequalities ........ . 299 6. Converse Inequalities ............................................ . 307 7. Generalizations of the Quasi-arithmetic Means ................... . 310 7.1 A Mean of Bajraktarevic ...................................... . 310 7.2 Further Results ............................................... . 316 7. 2 .1 Deviation Means .......................................... . 316 7.2.2 Essential Inequalities ...................................... . 317 7.2.3 Conjugate Means ......................................... . 320 7.2.4 Sensitivity of Means ....................................... . 320 CHAPTER V SYMMETRIC POLYNOMIAL MEANS ............. . 321 1. Elementary Symmetric Polynomials and Their Means ............. . 321 2. The Fundamental Inequalities .................................... . 324 3. Extensions of S(r;s) of Rado-Popoviciu Type ...................... . 334 v
X
4. The Inequalities of Marcus & Lopes .............................. . 5. Complete Symmetric Polynomial Means; Whiteley Means ......... . 5.1 The Complete Symmetric Polynomial Means .................. . 5.2 The Whiteley Means .......................................... . 5.3 Some Forms of Whiteley ...................................... . 5.4 Elementary Symmetric Polynomial Means as Mixed Means .... . 6. The Muirhead Means ............................................ . 7. Further Generalizations .......................................... . 7.1 The Hamy Means ............................................. . 7. 2 The Hayashi Means ........................................... . 7.3 The Biplanar Means .......................................... . 7.4 The Hypergeometric Mean .................................. . CHAPTER VI OTHER TOPICS .................................. . 1. Integral Means and Their Inequalities ............................ . 1.1 Generalities ................................................... . 1.2. Basic Theorems .. . .. . . .. . . . .. .. .... . .. . . . . . . . . .. . . . . ... . ..... 1.2.1 Jensen, Holder, Cauchy and Minkowski Inequalities ........ . 1.2.2 Mean Inequalities ......................................... . 1.3 Further Results ............................................... . 1.3.1 A General Result . . . . . .. . . . . .. . . . . . . . . . .. . . . . . .. . . . . . . . ... . 1.3.2 Beckenbach's Inequality; Beckenbach-Lorentz Inequality 1.3.3 Converse Inequalities ...................................... . 1.3.4 Ryff's Inequality .......................................... . 1.3.5 Best Possible Inequalities .................................. . 1.3.6 Other Results ............................................. . 2. Two Variable Means ............................................. . 2.1 The Generalized Logarithmic and Extended Means ........... . 2.1.1 The Generalized Logarithmic Means ....................... . 2.1.2 Weighted Logarithmic Means of n-tuples .................. . 2.1.3 The Extended Means ..................................... . 2.1.4 Heronian, Centroidal and Neo-Pythagorean Means ......... . 2.1.5 Some Means of Haruki and Rassias ........................ . 2.2 Mean Value Means ............................................ . 2.2.1 Lagrangian Means ........................................ . 2.2.2 Cauchy Means ............................................ . 2.3 Means and Graphs ............................................ . 2.3.1 Alignment Chart Means .................................. . 2.3.2 Functionally Related Means .............................. . 2.4 Taylor Remainder Means ..................................... . 2.5 Decomposition of Means ...................................... . 3. Compounding of Means .......................................... . 3.1 Compound means ............................................. . 3.2 The Arithmetico-geometric Mean and Variants. . ............... . 3.2.1 The Gaussian Iteration .................................... . 3.2.2 Other Iterations ........................................... . 4. Some General Approaches to Means
.
Xl
338 341 341 343 349 356 357 364 364 365 366 366 368 368 368 370 370 373 377
377 378 380 380 381 382 384 385 385 391 393 399 401 403 403 405 406 406 407 409 412 413 413 417 417 419 420
4.1 Level Surface Means . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ... . .... ........ . ..... ............ ... .... 4.2 Corresponding Means 4.3 A Mean of Galvani . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ... . . . .. . . . . . . . . . . . .. . ..... ... . .. 4.4 Admissible Means of Bauer 4.5 Segre Functions .............................................. . 4.6 Entropic Means ............................................... . 5. Mean Inequalities for Matrices ................................... . 6. Axiomatization of Means ......................................... . BIBLI 0 G RAPHY ..................................................... . Books . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Papers .............................................................. . NAME INDEX ....................................................... . INDEX ..................................... -......................... . ...
..
Xll
420 422
423 423 425 427
429 435 439 439
444 511 525
'Means and Their Inequalities" -Preface
There seems to be two types of books on inequalities. On the one hand there are treatises that attempt to cover all or most aspects of the subject, and where an attempt is made to give all results in their best possible form, together with either a full proof or a sketch of the proof together with references to where a full proof can be found. Such books, aimed at the professional pure and applied mathematician, are rare. The first such, that brought some order to this untidy field, is the classical "Inequalities" of Hardy, Littlewood & P6lya, published in 1934. Important as this outstanding work was and still is, it made no attempt at completeness; rather it consisted of the total knowledge of three front rank mathematicians in a field in which each had made fundamental contributions. Extensive as this combined knowledge was there were inevitably certain lacunre; some important results, such as Steffensen's inequality, were not mentioned at all; the works of certain schools of mathematicians were omitted, and many important ideas were not developed, appearing as exercises at the ends of chapters. The later book "Inequalities" by Beckenbach & Bellman, published in 1961, repairs many of these omissions. However this last book is far from a complete coverage of the field, either in depth or scope. A much more definitive work is the recent "Analytic Inequalities" by Mitrinovic, (with the assistance of Vasic), published in 1970, a work that is surprisingly complete considering the vast field covered. On the other hand there are many works aimed at students, or non-mathematicians. These introduce the reader to some particular section of the subject, giving a feel for inequalities and enabling the student to progress to the more advanced and detailed books mentioned above. Whereas the advanced books seem to exist only in English, there are excellent elementary books in several languages: "Analytic Inequalities" by Kazarinoff, "Geometric Inequalities" by Bottema, Djordjevic, Janie & Mitrinovic in English; "Nejednakosti" by Mitrinovic, "Sredine" by Mitrinovic & Vasic in Serbo-croatian, to mention just a few. Included in this group although slightly different are some books that list all the inequalities of a certain type-a sort of table of inequalities for reference. Several books by Mitrinovic are of this type. Due to the breadth of the field of inequalities, and the variety of applications, none of the above mentioned books are complete on all of the topics that they take up. Most inequalities depend on several parameters, and what is the most natural domain for these parameters is not necessarily obvious, and usually it is not the widest possible range in which the inequality holds. Thus the author, even Xlll
the most meticulous, is forced to choose; and what is omitted from the conditions of an inequality is often just what is needed for a particular application. What appears to be needed are works that pick some fairly restricted area from the vast subject of inequalities and treat it in depth. Such coherent parts of this discipline exist. As Hardy, Littlewood & P6lya showed, the subject of inequalities is not just a collection of results. However, no one seems to have written a treatise on some such limited but coherent area. The situation is different in the collection of elementary books; several deal with certain fairly closely defined areas, such as geometric inequalities, number theoretic inequalities, means. It is the last mentioned area of means that is the topic of this book. Means are basic to the whole subject of inequalities, and to many applications of inequalities to other fields. To take one example: the basic geometric mean-arithmetic mean inequality can be found lurking, often in an almost impenetrable disguise, behind inequalities in every area. The idea of a mean is used extensively in probability and statistics, in the summability of series and integrals, to mention just a few of the many applications of the subject. The object of this book is to provide as complete an account of the properties of means that occur in the theory of inequalities as is within the authors' competences. The origin of this work is to be found in the much more elementary "Sredine" mentioned above, which gives an elementary account of this topic. A full discussion will be given of the various means that occur in the current literature of inequalities, together with a history of the origin of the various inequalities connecting these means 1 . A complete catalogue of all important proofs of the basic results will be given as these indicate the many possible interpretations and applications that can be made. Also, all known inequalities involving means will be discussed. As is the nature of things, some omissions and errors will be made: it is hoped that any reader who notices any such will let the authors know, so that later editions can be more complete and accurate. An earlier version of this book was published in 1977 in Serbo-croatian under the title "Sredine i sa Njima Povezane Nejednakosti". The present work is a complete revision, and updating of that work. The authors thank Dr J. E. Pecaric of the University of Belgrade Faculty of Civil Engineering for his many suggestions and contributions. Vancouver & Belgrade 1988
1 Although not mentioned in this preface the book was devoted to discrete mean inequalities and did not discuss in any detail integral mean inequalities, matrix mean inequalities or mean inequalities in general abstract spaces. This bias will be followed in this book except in Chapter VI.
.
XIV
PREFACE TO THE HANDBOOK
Since the appearance of Means and Their Inequalities the deaths of two of the authors have occurred. The field of inequalities owes a great debt to Professor Mitrinovic and his collaborator for many years, Professor Vasic. Over a lifetime Professor Mitrinovic devoted himself to inequalities and to the promotion of the field. His journal, Univerzitet u Beogradu Publikacije Elektrotehnickog Fakulteta. Serija Matematika i Fizika, the "i Fizika' was dropped in the more recent issues, has in all of its volumes, from the first in the early fifties, devoted most of its space to inequalities. In addition his enthusiasm has resulted in a flowering of the study both by his students, P. M. Vasic, J. E. Pecaric to mention the most notable, and by many others. The uncertain situation in the former Yugoslavia has lead to many of the researchers situated in that country moving to institutions all over the world. There are now more journals devoted to inequalities, such as the Journal for Inequalities and Applications and Mathematical Inequalities and Applications, as well as many that devote a considerable portion of their pages to inequalities, such as the Journal of Mathematical Analysis and Applications; in addition mention must be made of the electronic Journal of Inequalities in Pure and Applied Mathematics based on the website http: I /rgmia. vu. edu. au and under the editorship of S. S. Dragomir. This website has in addition many monographs devoted to inequalities as well as a data base of inequalities, and mathematicians working in the field. Another welcome change has been the many contributions from Asian mathematicians. While there have always been results from Japan, in recent years there has been a considerable amount of work from China, Singapore, Malaysia and elsewhere in that region. It was taken for granted in the earlier Preface that anyone reading this book would not only be interested in inequalities but would be aware of their many applications. However it would not be out of place to emphasize this by quoting from a recent paper; [Guo & Qi]. "It is well known that the concepts of means and their inequalities not only are basic and important concepts in mathematics, (for example, some definitions of norms are often special means 2 ), and have explicit geometric 2
More precisely " ... certain means are related to norms and metrics.". See III 2.4, 2.5. 7 VI 2.2.1.
XV
meanings 3 , but also have applications in electrostatics4 , heat conduction and chemistry5 . Moreover, some applications to medicine 6 have been given recently." Due to the extensive nature of the revision in the second edition and the large amount of new material it has seemed advisable to alter the title but this handbook could not have been prepared except for the basic work done by my late colleagues and I only hope that it will meet the high standards that they set. In addition I want to thank my wife Georgina Bullen who has carefully proofread the non-mathematical parts of the manuscript, has suffered from computer deprivation while I monopolized the screen, and without whose support and help the book would have appeared much later and in a poorer form. P. S. Bullen Department of Mathematics University of British Columbia Vancouver BC Canada V6T 1Z2
[email protected] 3
See II 1.1, 1.3.1; [Qi & Luo].
4 See[P6lya f3 Szego 1951]. 5 See[ Walker, Lewis & McAdams], [Tettamanti, Sarkany, Kralik & Stomfai; Tettamanti & Stomfrul 6 See[ Ding].
.
XVI
BASIC REFERENCES
There are some books on inequalities to which frequent reference will be made and which will be given short designations. [AI] MITRINOVIC, D. S., WITH VASIC P. M. Analytic Inequalities, Springer-Verlag,
Berlin, 1970. [BB] BECKENBACH, E. F. & BELLMAN, R. Inequalities, Springer-Verlag, Berlin, 1961. [HLP] HARDY, G. H., LITTLEWOOD, J. E. & P6LYA, G. Inequalities, Cambridge University Press, Cambridge, 1934. [MI] BULLEN, P.S., MITRINOVIC, D. S & VASIC P.M. Means and Their Inequalities, D.Reidel. Dordrecht, 1988. [The first edition of this handbook.] [MO] MARSHALL, A. W. & OLKIN, I. Inequalities: Theory of Majorization and Its Applications, Academic Press, New York, 1979. Many inequalities can be placed in a more general setting. We do not follow that direction in this book but find the following an invaluable reference. Much of the material is readily translated to our simpler less abstract setting.
Convex Functions, Partial Orderings and Statistical Applications, Academic Press Inc., 1992. [PPT] PECARIC, J. E., PROSCHAN, F. & ToNG, Y. L.
There are two books that are referred to frequently in certain parts of the book and for which we also introduce short designations. [B 2 ] BORWEIN, J. M. & BoRWEIN, P. B.
Pi and the AGM. A Study in Analytic Number Theory and Computational Complexity, John Wiley and Sons, New York,1987. [RV] ROBERTS, A. W. & VARBERG, D. E. Convex Functions, Academic Press, New York-London, 1973. In addition there are the two following references. The first is a ready source of information on any inequality, and the second is in a sense a continuation of [AI] and [BB] above, being a report on recent developments in various areas of inequalities.
..
XVll
lDI] BuLLEN, P. S.
A Dictionary of Inequalities, Addison-Wesley Longman, Lon-
don, 1998. 7 [MPF] MITRINovr6, D. 8., PECARIC, J. E. & FINK, A. M.
Classical and New In-
equalities in Analysis, D Reidel, Dordrecht, 1993. There are many other books on inequalities and many books that contain important and useful sections on inequalities. These are listed in Bibliography Books. From time to time conferences devoted to inequalities have published their proceedings. In particular, there are the proceedings of three symposia held in the United States, and of seven international conferences held at Oberwolfach.
Inequalities, Inequalities II, Inequalities III, Proceedings of the First, Second and Third Symposia on Inequalities, 1965, 1967, 1969; Shisha, 0., editor, Academic Press, New York, 1967, 1970, 1972.
11(1965), 12 (1967), 13 (1969)
GI1(1976), GI2 (1978), GI3 (1981), GI4 (1984), GI5 (1986), GI6 (1990), GI7 (1995)
General Inequalities Volumes 1-7, Proceedings of the First-Seventh International Conferences on General Inequalities, Oberwolfach, 1976, 1978, 1981, 1984, 1986, 1990, 1995; Beckenbach, E. F., Walter, W., Bandle, C., Everitt, W. N., Losonczi, L., [Eds.], International Series of Numerical Mathematics, 41, 47, 64, Birkhaiiser Verlag, Basel, 1978, 1980, 1983, 1986,1987, 1992, 1997. Individual papers in these proceedings, referred to in the text, are listed under the various authors with above shortened references. Finally there are two general references. [EM1], [EM2], [EM3], [EM4], [EM5], [EM7], (EM8], [EM9],(EM10], [EMSuPPl], (EMSuPP II], [EMSuPP III]; HAZELWINKEL, M., [Eo.] Encyclopedia of Mathematics,
vol.1-10, suppl. I-III, Kluwer Academic Press, Dordrecht, 1988-2001. [CE] WEISSTEIN, E. W. CRC Concise Encyclopedia of Mathematics, Chapman & Hall/CRC, Boca Raton, 1998.
7 Additions and corrections can be found at http:/ /rgmia.vu.edu.au/monographs/bullen.html.
XVlll
NOTATIONS
1 Referencing Theorems, definitions, lemmas, corollaries are numbered consecutively in each section; the same is true of formulre. Remarks and Examples are numbered, using Roman numerals, consecutively in each subsection, and in each sub-subsection. In the same chapter references list the section, (subsection, sub-section, in the case of remarks and examples), followed by the detail: thus 3 Theorem 2, 4(6), or 1.2 Remark (6). Footnotes are numbered consecutively in each chapter and so are referred to by number in that chapter: thus 1.3.1 Footnote 1. References to other chapters are as above but just add the chapter number; thus I 3 Theorem 2(a), II 4(6), IV 1.2 Remark (6), I 1.3.1 Footnote 1. Although there are references for all names and all bibliography entries, in the case of a name occurring frequently, for instance Cauchy, only the most important instances will be mentioned; further names in titles of the basic references are usually not referenced, thus Hardy in [HLP].
2 Bibliographic References Some have been given a shortened form; see Basic References. Others are standard, the name, followed by a year if there is ambiguity, or the year with an additional letter, such as 1978a, if there is more than one any given year. Joint authorship is referred to by using &, thus Mitrinovic & Vasic.
3 Symbols for Some Important Inequalities Certain inequalities are referred to by a symbol as they occur frequently. (B) ...................................... Bernoulli's inequality I 2.1 Theorem 1; (C) ................................................. Cauchy's inequality III 2.2; (GA) .................. Geometric-Arithmetic Mean inequality II 2.1 Theorem 1; (H) ....................................... Holder's inequality III 2.1 Theorem 1; (HA) .................. Harmonic- Arithmetic Mean Inequality II 2.1 Remark(i); (J) ........................................ Jensen's inequality I 4.2 Theorem 12; (M) .................................. Minkowski 's inequality III 2.4 Theorem 9; (P) ..................................... Popviciu's inequality II 3.1 Theorem 1; (R) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Rado 's inequality II 3.1 Theorem 1; (r;s) ................................ Power Mean inequality III 3 .1.1 Theorem 1; S(r;s) .... Elementary Symmetric Polynomial Mean inequality V 2 Theorem 3(b);
.
XIX
(T), (TN) ........................................... Triangle inequality III 2.4 . Integral analogues of these means, when they exist, will be written (J)- J, etc; see VI 1.2.1, 1.2.2.; and (rvB), etc. will denote the opposite inequalities; see 9 below.
4 Numbers, Sets and Set Functions Z, Q, JR, C are standard notations for the sets of integers, rational numbers, real numbers, and complex numbers respectively. The set of extended real numbers, ~ U { ±oo}, is written JR. Less standard are the following: N == {n; n E Z, and n > 0} == {0, 1, 2, ... } , N* == N \ {0} == { 1, 2, ... } , ~ * == JR \ { 0}, ~+ == {X; X E JR and X > 0}, ~+ == JR.+ \ { 0} == {X; X E JR. and X >
0}, Q* == Q\ {0}, Q+ == {x;x E Qandx > 0},
Q+ == Q+ \ {0} == {x;x E Qandx >
0}. The non-empty subsets of N, or N*, are called index sets, the collection of these is written T. If p E JR*, p =/= 1, the conjugate index of p, written p', is defined by
'
p p == p- 1'
equivalently
1
1
-p +== 1. p'
Note that
(p')' == p,
p > 1 ~ p' > 1,
p > 0 and p' > 0 ~ p > 1, or p' > 1.
A real function ¢ defined on sets, a set function, is said to be super-additive, respectively log-super-additive, if for any two non-intersecting sets, I, J say, in its domain
¢(I U J) >¢(I)+ ¢(J), respectively, ¢(I U J) > ¢(I)¢(J). If these inequalities are reversed the function is said to be sub-additive, respectively log-sub-additive. A set function f is said to be increasing if I C J ~ f (I) < f (J), and if the reverse inequality holds it is said to be decreasing. The image of a set A by a function f will be written f[A]; that is f[A] == {y; y ==
f(x), x E A}. A set E is called a neighbourhood of a point x if for some open interval ]a, b[ we have x E [a, b[C E.
5 Intervals Intervals in JR. are written [a, b], closed; ]a, b[, open etc. In addition we have the unbounded intervals [a, oo[, ] - oo, a], etc; and of course]- oo, oo[== JR., [-oo, oo] == JR., [O,oo[== JR+, ]O,oo[== JR+. The symbol (a,b) is reserved for 2-tuples, see the 0
next paragraph. If I is any interval then I denotes the open interval with the same end-points as I.
XX
6 n-tuples If ai E ~' CC, 1 < i < n, then we write a for the, ordered, real, complex, n-tuple (a 1, ... , an) with these elements or entries; so if a is a real n-tuple a E ~n. The usual vector notation is followed. In addition the following conventions are used. (i) For suitable functions f, g etc, f(a) = (f(al), ... , f(an)), and g(a, b) = (g(a 1, b1), ... , g(an, bn)) etc. This useful convention conflicts with the standard notation for functions of several variables, f(a) = f(al, ... , an), but the context will make clear which is being used. So: a 2 = (ai, . .. ,a;), ab = (a1b1, ... ,anbn); but maxa = maxl 0 etc, we say that the n-tuple is positive, non-negative etc. The set of non-negative n-tuples is written ~+, and the set of positive n-tuples (JR~)n
(iii) An n-tuple all of whose elements are all zero except the i-th that is equal to 1 is written ei. The n-tuples ei, 1 < i < n, form the standard basis of 1Rn. As usual if n = 2, 3 these basis vectors are written e 1 = i, e 2 = j, ~ = k. (iv) e is then-tuple each of whose entries is equal to 1 ; and 0, or just 0, is . the n-tuple each of whose entries is equal to 0. (v) If 1 < i < n then a~ denotes the (n - 1)-tuple obtained from a by omitting the element ai. (vi) a rv b, a is proportional to b, means that for some A., J.-L E JR, not both zero, A.a + J.-Lb = 0; that is the two n-tuples are linearly dependent. (vii) If ai = c, 1 < i < n, we say that a is constant, or is a constant. (viii) For D._kan and see I 3.1. Ann-tuple is called an arithmetic progression if 6. 1 ak, 1 < k < n- 1, is constant, equivalently if D.. 2 ak = 0, 1 < k < n- 2, (ix) For a* b see I 3.2 Definition 4. (x) For a-< b see I 3.3 Definition 11. wi, 1 < i < n; and if (xi) Given an n-tuple w we will write Wk = 1 necessary W 0 = 0. More generally if w is a sequence and I E I, an index set, see 4, we write W1 = 'L:iEJ Wi· (xii) If f is a function of k variables, a and n-tuple, 1 < k < n, then L k ff (ai 1 , ••• , aik) means that the summation is taken over all permutations of k elements from the a1, ... , an. In the case that k = n this will be written as 'L:fJ(a). (xiii) The inner product of two n-tuples, a, b, written (a, b), is 'L:~ 1 aibi if the n-tuples are real and 'L:~ 1 aibi if the n-tuples are complex. In both cases (a, a)= I:~ 1 Where appropriate the above notations will be used for sequences a = ( a1, a2, ... ) , provided the relevant series converge.
L::
a;.
.
XXI
7 Matrices An m x n matrix is A == ( aij) l~i~rn ; the aij are the entries and if they are real, l~j~n
complex then A is said to be real, complex. The transpose of A is AT== (aji) l~i~rn;
A*
l<j ( ai 1 , . . . , ain) E D for every permutation (i1, ... , in) of (1, ... , n), (iii) f(ai 11 ••• , ain) == f(al, ... , an) for every permutation (i 1 , . . . , in) of (1, ... , n), then f is said to be symmetric. A related concept is that of an almost symmetric function: this is a function of n variables and n parameters, defined on a symmetric domain, that is invariant under the simultaneous permutation of the variables and the parameters. EXAMPLES(i) The arithmetic mean with equal weights, f(a) == (a 1 + · · ·+an)/n, is symmetric; the weighted arithmetic mean, f (a) == f (a; w) == (w1a1 + · · · + wnan)/(wl + · · · + wn), is almost symmetric; see II 1.1. The Gamma or factorial function is denoted by x!; that is
x! == r(x + 1). The identity function is written
. lS
i,
and then the power functions
i8
,
s E JR*. That
i(x) == x; 8 This conflicts with the use of I to denote an interval but in practice there will be no confusion .
..
XXII
the domain being clear from the context. The maximum function is defined as usual by max{f, g}(x) == max{f(x), g(x)}: also,
x+ ==max{ x, 0}; x- ==max{ -x, 0}; x == x+ - x-, lxl == x+
+ x-.
By analogy iff is any real-valued function then j+ == max{f,O} == maxof; when f == j+ - f-' lfl == j+ +f-. The function that is equal to 1 on the set A and to zero off A, the indicator function 9 of A is written 1A. That is if X E A, if X~ A. The integral or integer part function [·] : R
~
Z is defined by:
[x] == n, where n E Z is the unique integer such that n
.....................
L(X) , t 8 (X) . . . . . . . . . . . . . . . . . . . . . . . xxii x+, x-, j+
................... xxn1 ........................ xxiii
lA(x) [X] ........................... XXlll f ~ (1fo ; v) . . . . . . . . . . . . . . . . . . . . . . . 51
f[A] ........................... XX [a, b], ]a, b[, etc .................. xx [a,oo[,]-oo,a]
[~]
I
.............................. XX f (a) , g (a , b) . . . . . . . . . . . . . . . . . . . . xxi max a, m1n a ................... xxi a ..x, A.y) == A.f(x, y), and monotonicity, x < x', y < y' ====> f(x, y) < f(x', y'); finally there is the crucial property of internality that justifies the very name of . mean, min{x, y} < f(x, y) < max{x, y}. Most of the means introduced are easily defined for n-tuples, n > 2, when these various properties are suitably extended; see II 1.1 Theorem 2, 1.2 Theorem 6, III Theor~m 2(e) and VI 6. Of course there are many other properties of means that have been identified as of interest; these are listed in the Index under Mean Properties. The inequalities between the various means defined form the core material of the book. Again the two variable cases are the easiest II 2.2.1, 2.2.2, 5.5, VI 2, 3.1, 3.2.1. The fundamental result is the inequality between the arithmetic and geometric means, (GA), discussed in detail in II 2.4 where well over 70 proofs are given or mentioned; most are extremely elementary. The next basic result generalizes this and is the inequality between the power means, (r;s), III 3.1.1. Integral forms of these results are also give; VI 1.2.2. From Notations 9 we see that every inequality between means of n-tuples can be regarded as saying a certain function of n is non-negative. For instance ( G A) im-
.
XXVI
plies that R(n) == n(2tn- ®n) > 0. A stronger property of this function of n would be to show that it increases; stronger because R(1) == 0; a similar discussion occurs for related functions that are not less than 1, (2tn/®n)n > 1 for example. This leads to the so called Rado-Popoviciu type extensions of the original inequality. Such are discussed for (GA) in II 3.1; the analogous discussion for (r;s) in III 3.2 is much more technical as the simplicity of the (GA) case has been lost. Finally many well-known inequalities arise from the discussion of mean inequalitiesin particular the inequalities of Cauchy, {C), Holder, (H), A1inkowski, {M}, Cebisev and the triangle inequality, (T); see II 5.3, III 2.1, 2.2, 2.4 .
..
XXVll
I
INTRODUCTIO N
In this chapter we will collect some results and concepts used in the main body of the text. There is no intention of being exhaustive in any of the topics discussed and often the reader will be referred to other sources for proofs and full details.
1 Properties of Polynomials Simple properties of polynomials can be used to deduce some of the basic inequalities to be discussed in this book. In addition certain simple inequalities, needed at various places, are most easily deduced from the properties of some special polynomials. These results are collected together in this section for ease of reference. CONVENTION
In t his sect ion , u n 1 e s s o t her wise s p e c i fie d , a 11
polynomials will have real coefficients. 1.1 SOME BASIC RESULTS
The results given here are standard and proofs are
easily available in the literature; see for instance [CE pp.420, 1573; DIp. 70; EM3
p.59; EMB p.175], [ Uspensky]. THEOREM
1 A polynomial of degree n has n complex zeros, and if n is odd
at
least
one zero is real. A polynomial cannot have more positive zeros than there are variations of signs in its sequence of coefficients. THEOREM
2
[DESCARTES' RuLE OF SIGNs]
If p is a polynomial then p' has between any two distinct real zeros of p.
THEOREM
THEOREM
3
4
[RoLLE's THEOREM]
least one zero
A polynomial always has a zero between which its values are of opposite sign.
[INTERMEDIATE VALUE THEOREM]
any two numbers THEOREM
at
at
5 A zero of a polynomial is
a
zero of its derivative if and only if it is
multiple zero. 1
a
Chapter I
2
The following result is basic to a variety of applications; see [BB p.11; HLP pp.1041 05], [Milovanovic, Mitrinovic €3 Rassias pp. 70-71; Newton], [Dunkel 1908/9; Kellogg; Maclaurin; Sylvester]. The present form is that given in [HLP pp. 104105 ]. THEOREM
If f(x,y) ==I:~ oCiXiyn-i has, as a function ofyjx, all of its zeros
6
real then the same is true of all polynomials, not all of whose coefficients are zero, derived from f by partial differentiations with respect to x or y. Further if a zero of one of these derived polynomials has multiplicity k, k multiplicity k
0
+ 1 of the polynomial from
>
1, then it is a zero of
which it was obtained by differentiation.
The proof is immediate by repeated applications of Theorems 1, 3, 4 and 5.
0 7 If n
COROLLARY
> 2, and p(x) =
~ e;xi = ~ (7)dixi
(1)
is a polynomial of degree n with c0 =/= 0 and all zeros real, then if 1
< m < k+m < n
the polynomial q(x) ==I:~ 0 (7)dk+ixi has all of its zeros real.
0
Let f(x, y) ==I:~
is not a zero of
f.
0
(7)dixiyn-i. We are given that do =/= 0, so (0, y), y =/= 0,
Hence, by Theorem 6, (0, y), y =1- 0, is not a multiple zero of
any derived polynomial. This implies that no two consecutive coefficients of
f can
vanish.
I
Save for a numerical factor an-m f akxan-k-my is equal to L~ 0 (7)dk+ixiym-i' which by the previous remark does not have all of its coefficients zero. Hence the D
result is an immediate consequence of Theorem 6. REMARK
It follows from the above proof that if pis a polynomial of degree n,
(i)
as in (1), with co =1- 0, and if for some k, 2
< k < n- 1, ck
== ck-1 == 0, then p has
a complex, non-real, zero; [Wagner C ]. CoROLLARY
8
If n
> 2, and p a polynomial of degree n given by (1) with co =1- 0
and all of its zeros real, then fork, 1
< k < n- 1,
d~ >dk-1dk+1,
(2)
2
(3)
ck >ck-1Ck+1· The inequalities (2) are strict unless all the zeros are equal.
0
By Corollary 7 the roots of all the equations dk-1
+ 2dkx + dk+1x 2 == 0,
1
< k < n- 1,
(4)
Means and Their Inequalities
3
are real; from this (2) is immediate. Now from (2) and the definition of dk, 1
unless possibly either
Ck-1
< k < n, see (1),
== 0, or ck+1 == 0; but then, by Remark (i), Ck
# 0;
and
so in all cases (3) is proved. Finally, if for any k there is equality in (2) the associated quadratic equation (4) has a double root, and so, by Theorem 6, the original polynomial p of (1), has a single zero of multiplicity n. REMARK
(ii)
D
Inequality (2) is sometimes called Newton's inequality and will reap-
pear later, II 2.4 proof ( ix), V 2 Theorem 1; [D I p .1 92 J. A direct proof can be found in [Nowicki 2001]. A converse to this inequality has been given; [Whiteley 1969]. REMARK
(iii)
The above implies that if for some k, 1 < k
< n -1,
c~
< Ck-1Ck+1,
then the above polynomial p has a non-real zero. REMARK
(iv)
By writing (2) in the form d~k
>
d~_ 1 d~+ 1 , 1
-1, x =I 0, and if 0 < a < 1 then
+ x )a
1 or a < 0, when of course x
0
1 +ax; =/=-
(B)
-1, then (rv B) holds.
(i)O0, in particular iff' >O with f' (x )=0 at only a finite number of points, then f is strictly increasing.
of differentiation can be used to prove that: if
7
Means and Their Inequalities
x == a, say. If a == e, a == a , if 1 < a < e then a > e, while if a > e, then 0
1 < a < e; for a < 1 take a == oo. If I == [a, a], or [a, a], then ax < xa if x EI and ax
> xa if x
If x
~I; [Bullen 2000; Qi & Debnath 2000b; Smirnov].
# 0 then ex> 1 + x.
(8)
This follows in several ways: ( i) the strict convexity of the function ex implies by 4.1 Corollary 3 that at each point its graph lies above the tangent, and y
== x + 1
== 0;
is the tangent to the graph at x
x2
( ii) by Taylor's theorem since ex == 1 + x + eY, for some y between 0 and x; 2 (iii) note that f (x) == ex - 1 - x has a unique maximum at x == 0. 2
REMARK
(ii)
Of course the Taylor's theorem argument can be used to prove that
xn ex > 1 + x + · · · + -, , x > 0. In particular, taking n == 2 and replacing x by x- 1 n. 1 we get ex-l jx > (x + 1/x), x > 1.
2 Another well-known fact is that 1 )n ( 1+-
1 )n+l , <e< (1+-
n
n EN*.
n
It can be shown that the left-hand side increases to e, and the right-hand side decreases to e, as n ----+ oo; see II 2.5.3( 6). A simple deduction from these inequalities is a crude estimate for the factorial function n!, n EN*; [Klambauer p.410].
More generally, [DI p.81], [Wang C L 1989a], if x ex
(1 + -n1)-x < (1 + -nX)n < ex,
which shows that limn---+oo ( 1 + (b)
# 0,
THE LOGARITHMIC FUNCTION
n EN*,
x) n == ex.
n
[DI pp.J59-J60] If X> 0, X-/= 1 then
x-1
- - < log X <X- 1.
(9)
X
2
Taylor's theorem states:if f
has n, n>O, continuous derivatives on [a,b], and if t 0 and put
f(x)
a
b
+ -, x > 0. x
1 2 - ab) we see that f has a unique minimum at x == (x ax 2 VOJ) and that f(xo) ==Yo == 2~. In particular f is strictly convex, strictly
Noting that f'(x)
x 0 ==
X
==-
== -
decreasing on the interval ]0, xo], and strictly increasing on the interval [x 0 , oo[. Alternatively the equation f(x) == y has two roots, x1, x2 say, and x1x2 == ab, x1
+ x2
== ay. So
If
X b -+->2 -,
a
x
(17)
a
with equality if and only if x == VOJ); further if x
> x 3 > VOJ) then b X3 b -+->-+a x - a X3 ' X
with equality if and only if x == and finally if x 1
(18)
X3;
< x < x 2 then 2
X
b x
X1
b x1
X2
b x2
0, or by noting that (x- 1)(1- 1/x) > 0. (ii) If a= _!_ a x > 1 then
< 1,
b = 1 then x 0 =
~ < 1 So taking X3 =
1 in (18) we get that if
ya
1
(21)
ax+-> a+ 1, X
with equality if and only if x == 1. This is an inequality due to Chong, [Chong K
M 1983], A direct proof can be given as follows: 1
ax+- ==(a- 1)x + (x X
1
+ -) X
>(a- 1)x + 2, by (20), >a+ 1. For equality at the first step we need x == 1, and then there is equality at the second step. Inequality (21) can be written in an equivalent form: put a == ujv, u > v > 0 then v ux + X
> u + v,
(22)
with equality if and only if x == 1. (iii) Given m, M, 0 < m < M choose x 1 == m, x2 == M and a == b == vrn:Nf then from (19) 2
x vrn:Nf < < + - vr;:M x -
fl-+ ~ m
m+M -== --M
with equality in the left inequality if and only if x inequality if and only if x == m or x
vr;:M '
(23)
== vrn:Nf, and in the right
== M.
Inequality (20) has been generalized by Korovkin; [Korovkin p.B]. LEMMA
6 If x is a positive n-tuple, n
> 2, then (24)
Means and Their Inequalities
11
with equality if and only if x is constant. 0
The proof is by induction, noting that the case n
a == b == x 2 and x Xn
== x1.
==
2 is just (20) with
So assume (24) holds for integers less than n, and that
==minx. Then the left-hand side of (24) is equal to X1 -+···+ X2
Xn-1 X1
> ( n- 1) +
+
==n
Xn Xn-1) + (Xn-1 + -X1Xn X1 Xn-1
(
Xn
Xn Xn-1) +-, X1
by the induction hypothesis,
X1
x1)
(xn- Xn-1)(xnX1Xn
> n.
The case of equality is immediate. REMARK
D
For further generalizations see II 2.4.5 Lemma 19 and II 2.5.3 (¢)
(vi)
Theorem 25.
3 Properties of Sequences 3.1 CONVEXITY AND BOUNDED VARIATION OF SEQUENCES
Let a
== (a1, a 2 , ... )
be a real sequence and define the sequences ~ k a, k E N*, by recursion as follows: 3
n EN*·
'
Then it easily follows that
(1)
and that if 1
< j < k,
.
6..1 an CONVENTION
3
=
~
~
(i +
Through o u t
We will also use the notation So .6_kan=(-l)kAan, k,nEN*.
~ka,
where
k - j -
k_j _ 1
t his
2) ~
k
(2)
ai+n-1·
sect ion
we
assume
t hat :
~an=-Aan=an+l-an, ~kan=~(~k-lan),
nEN*, k>2.
12
Chapter I
(a) A sequence a is said to be k-convex, k is non-negative. DEFINITION 1
E
N*, if the sequence ~ k a
(b) A sequence a is said to be of bounded k- variation, k
E
N*, if
A sequence that is 1-convex is exactly a decreasing sequence; a 2convex sequence, usually called a convex sequence, has an- 2an+1 + an+2 > 0, n > (i)
ExAMPLE
1.
A sequence a is of 1-bounded variation, or just bounded variation, if :L~ 1 lai - ai+11 < oo. Clearly the sequence of partial sums of a series is of bounded variation if and only if the series is absolutely convergent. (ii)
EXAMPLE
Iff is a convex function, see 4.1 Definition 1, then an == f(n), n 1, . . . is a convex sequence; more generally if f is k-convex, see 4. 7, then an ( -l)n f(n), n == 1, ... is k-convex. [DI pp.64-65, 191, EM2 p.419].
REMARK
(i)
REMARK
(ii)
== ==
If the sequence -a is convex, k-convex we say that a is concave,
k-concave. REMARK
(iii)
The definition of a k-convex n-tuple, n > k, is immediate.
The partial sums of a series can be written as the difference of the partial sums of
2:7
2:7
2::7
two positive termed series, Bn == 1 b-:};1 bn == 1 bn == Pn - Nn; and the series L bn is absolutely convergent if and only if both of the series 1 b-:}; and 1 bn converge. This can be expressed in terms of the present concepts as:
:2:7
2::7
if a sequence is bounded then it is of bounded variation if and only if it can be written as the difference of two bounded decreasing sequences. The main result of this section, Theorem 3 below, is a generalization of this result due to Dawson, [Dawson].
First we prove a basic lemma
If a is a sequence that is bounded and k-convex, respectively of bounded k-variation, k > 2, then: (a) a is p-convex, respectively bounded p-variation, 1 < p < k - 1;
LEMMA
(b) (c)
2
. (n + 1) . (i + 2) · L
hm
n-+oo
oo
.
'l=l
j. J
.j -
J- 1
f~_J an
.
==0, 1 < J < k - 1;
~Jai==a1-
lim an, l<j 0. This is (a) in this case. Now
n
n-1
L flai L ill ai + n!lan,
a1- an+1 ==
2
==
(3)
i=1
i=1
and noting that !la > 0, fl 2 a > 0, and that a is bounded, the series in (3) converges, which implies (c). Hence limn~oo an exists and is finite, which from (3) and the convergence of the series shows that limn~oo nil an exists and is non-negative, and so is zero. This gives (b) and completes the proof of this case. ( ii): k == 2 and a is bounded and of bounded 2-variation. From (3) the sequence n!lan, n EN*, is bounded and hence I:~ 1 !flail converges, which is (a) in this case. Further limn~oo an exists, and so from (3) limn~oo n!lan exists, A say, and assume that A =I= 0. Then for some no E N*,
IAI 2
D.
00
.L
1-
~:: < L il~ail
D.
00
1
1-
~::
1
~=no
~=no
00
==
L
2
ijfl ail < oo.
i=no
Hence
rr
oo ( flai+ 1 ) fla·
.
.
converges absolutely to some non-zero hmit. This contradicts
~
~=no
the convergence of~~ 1 j!lail, and so A== 0. The lemma in this case now follows from (3). (iii): k > 2 and a is bounded and k-convex. Since !lka == ll 2 (D.k- 2a) the sequence D.k- 2a is convex and bounded. Hence (a),
in this case, follows by induction. In particular a is convex and so (b) and (c) hold with j == 1, and j
== 1, 2 respec-
tively. Suppose that 1 < j 0 < k and that (c) holds for j
== Jo.
Then since
+ (n + Jo- 1)/ljo (i + Jo- 2)~Jo . ~ L...-t (i + Jo. -1)flio+ Jo Jo Jo + Jo - 1) · (n .. . ~ < L...-t (i + Jo. - 1) · Jo Jo .
~ ~
. _
i=1
1
a~
1
=
~=1
A say; assume that A =I= 0.
.
an,
(4)
i=1
flJo+ 1 ai
this 1mphes
. a~
oo, and lim
n~oo
.
flJo+lan exists,
Chapter I
14 Since
(n+ Joj 1) ~jo 0 .
an
== (
I:n .
~=1
(i +Jo jo- - 2)) ~jo a l
.
it follows that there is an n 0 E N* such that if n
. . Th1s contrad1cts the convergence of
00
~
~
('
2 and a is bounded and of bounded k-variation. Suppose that no subsequence of + k- 2) ~k-lan, n E N* converges to zero.
(n k-1
Then for some no
E
N* and A > 0 we have that
00
O,y > 0.
y)
(13)
r=O
THEOREM
8 (a) If If a and b are log-concave, so is a* b.
(b) If If a is a-log-convex, and b is (1 -a)-log-convex, 0 < a < 1, then a * b is log-convex. D
(a) [Menon 1969a] If c ==a* b then, n
n+1
n-1
c;- Cn-1Cn+1 ==(Larbn-r)
2
(Larbn-r-1)(Larbn-r+1) r=O r=O
-
r=O n-1
n
n-1
n
r=O
r=O
r=O
r=O
n
+ anbo ( L
n-1
arbn-r) - an+lbo ( L arbn-r-1) r=O r=O
== A+B + C,
Chapter I
20
say, where
A==
n-1 n
n-1 n
r=Ok=l
r=0k=1
L L arak(bn-rbn-k- bn-r-lbn-k+l) == L L dr,k, say,
n-1
B
==
L arao(bn-rbn- bn-r-1bn+1), r=O
C ==anbo
n
n-1
r=O
r=O
L arbn-r- an+1bo L arbn-r-1 n
+ bo L
==anbnaobo
bn-r (anar - an+1ar-1) ·
r=1
B and C are easily seen to be non-negative by the log-concavity of a and b. Since, in A, dr,r+1 == 1 we get on combining dr-1,k and dk-1,r that, n
A==
L (dr-1,k + dk-1,r) r,k=l r 0, and so the inequalities
(v)
REMARK
(g), for c, are strict. In part (b) however these inequalities can be equalities, and are so if and only if all the inequalities for a, and for b are equalities. (vi)
REMARK
The results in (b) is best possible in that the analogue of (a) is in
general false; further the result does not hold in general if a == 1. The following examples illustrate this.
a, {3 be two sequences defined as in (11); then a* {3 ==a+ {3 and (v) If -so if a+/3 > 1 the convolution is not log-convex but only a'-log-convex, a' > a+f3. EXAMPLE
EXAMPLE
1 Let an == -, bn == 1, n E N*, then a is 0-log-convex, see Example
(vi)
n
(ii), and b is log-convex; but a* b == c where
Cn
==
:z=;
1
1/r, n
E
N*, which is not
log-convex. 3.3 AN ORDER RELATION FOR SEQUENCES
Ann x n matrix is S == (sij)l 0. (e) Iff and g are convex, increasing {decreasing) and non-negative then fg is also convex. (f) Iff is convex on I and g is convex in J, f[I] C J, g increasing, then go f is convex on I. See [RV PP-4-7,15-16]. 5
0
REMARK
(viii)
If
D
f is an affine function then (f) holds without the assumption
that g is increasing. COROLLARY
5 (a) f is convex, respectively strictly convex, on ]a,
b[ if and only if
== f(xo) + m(x- xo) such respectively S(x) < f(x), a< x < b, x #- xo.
for each x 0 , a < xo < b, there is an affine function S(x)
that S(x) < f(x), a< x < b, (b) If f is differentiable and strictly convex on an interval I and if for c E I we have f'(c) == 0 then f(c) is the minimum value off on I and this minimum is
.
un1que D
(a) Iff is convex, respectively strictly convex, take inS any value of m such
that mE [f~(xo),f~(xo)]. Conversely suppose that such an S exists at xo and let xo
== AX+ (1 - A)y where x, y E]a, b[, 0 < A< 1. Then f(xo) == S(xo) == AS(x) + (1- A)S(y) < Aj(x) + (1A) f (y), showing that f is convex. The strictly convex case is similar. (b) This is an immediate consequence of Theorem 4(b) and Corollary 3(a); see
[Tikhomirov, p.117]. REMARK
(ix)
D
The affine function S is called a support off at xo; its graph is a
line of support for f at xo; [RV p.12]. 4
The Cantor function is continuous, increases from 0 to 1, and is constant on the dense set of intervals [1/3,2/3],[1/9,2/9],[7/9,8/9], ... , where it takes the values 1/2,1/4,3/4, ... ; see [CE p.187; EM2 p.13; P6lya & Szego p.206]. A set, or a set of intervals, is dense if every neighbourhood of every point meets the set, or the set of intervals; [CE p.416; EM3 PP-46,434]. 5 A function f is Lipschitz if for some M>O and all x,y, lf(y)-f(x)I<Miy-xl; then Miscalled a Lipschitz constant of the function; [OE p.1091; EM5 p.532].
Chapter I
28 THEOREM
6 (a)
f:
[a, b]
~
R is (strictly) convex if and only if there is a (strictly)
increasing function g :]a, b[~ R and a c, a
f(x)
=
< c < b, such that for all a, a < x < b,
f(c)
+
1x
g.
(b) Iff" exists on ]a, b[ then f is convex if and only iff"
> 0; if every subinterval
contains a point where f" > 0 then f is strictly convex. 0
0
See [RV pp.9-11].
REMARK
(x)
For applications to inequalities the conditions of (b) usually suf-
fice; that is we are dealing with functions that are twice differentiable, with second derivative positive on a dense set 6 -
usually the complement of a finite
set; see [Bullen 1998]. It might be noted that, using the mean-value theorem of differentiation 7 , we can, deduce from the last property that the chord slope increases strictly; see Lemma 2. COROLLARY
ifO
7 (a) If r
> 1 orr < 0 and f(x)
= xr, x
> 0, then f is strictly convex;
< r < 1, f is strictly concave. The exponential function is strictly convex and
the logarithmic function is strictly concave. (b) If f E C2 (a, b) with m = min f", M = max f" then both of the functions Mx 2 mx 2 are convex. - f(x) and mf(x)2
0
2
D
Simple applications of Theorem 6(b).
ExAMPLE
(i)
Other examples are: x log x, x > 0 is strictly convex; iff is (strictly)
convex, twice differentiable on I and if f < 1 then 1/ (1- f) is also (strictly) convex on I. These simple examples and those in Corollary 7(a) are used in many places to prove various inequalities; [Abou-Tair & Sulaiman 1999]. ExAMPLE
(ii)
The factorial function x!, x
>
-1, is strictly convex but more is
true in this case; see 4.5.2 Example (i). EXAMPLE
xr
(iii)
From Corollaries 7(a) and 3(a) we have: if x
> 0, x =/= 1, then
], 1 + r(x- 1), if 0 < r < 1, [r > 1, orr < OJ. This, with a simple change
of variable and notation is just 2.1 Theorem 1, that is (B), [( rvB)] . REMARK
(xi)
Corollary 7(b) has been used to extend many of the inequalities
implied by convexity; [Andrica & Ra§a; Ra§a]. 6 See Footnote 4 7 See Footnote 1.
Means and Their Inequalities
29
A simple property of convexity and concavity, that we will use later, is given in the following lemma. LEMMA 8
Iff : [0, 1]
[KuANG]
~
1R is strictly increasing and either convex
or
concave, tben for all n > 1,
f is strictly increasing. Suppose then that f is convex and that 1 < i < n + 1, then 0
The last inequality holds since
i~1~c~1)+( 1 - i~1)!(!) > f ( ( i - 1)2 +(1n
1 ) i) ' n n
i -
-n;) + 1)
=JC(n >f ( n :
1
) , since
by ( 1)'
f is strictly increasing and i < n + 1.
Summing we get that
1
n
.
1
.
n
.
-L((i-1)t(~- )+(n-i+1)t(~))>Lt( ~ ). n. n n . n+1 ~=1
~=1
So n-1
.
-n.L ((i- 1)t(~-n 1
1
.
1
) + (n- i + 1)t( ~ )) + -t(1) > n
~=1
n
n+1
.
L. t( n+1 ~ ) - t(1). ~=1
This on simplification gives the first inequality. A similar argument, starting with
to the same inequality when REMARK
(xii)
1 i f ( i + ) + (1 i )f ( i ) n+ 1 n+1 n+1 n+1
f is concave.
leads D
This result is a special case of a slightly more general result;
[Kuang; Qi 2000a]. Another useful result is the following; see [DI pp.122-123]. THEOREM
9
[HADAMARD-HERMITE INEQUALITY]
Iff : [a, b]
~
1R is convex and if
a < c < d < b, tben c+d)< 1 1df 0, aP
-<w bP.
0 We can without loss in generality assume that the terms in the n-tuples are all greater than 1. Then apply Theorem lO(b)with function f(x) = ePx and the n-tuples log a, log b. D 4.2 JENSEN's INEQUALITY
One of the most important convex function inequal-
ities is Jensen's inequality; in fact it is almost no exaggeration to say that all 8 It is sometimes called the Hermite-Hadamard inequality
Means and Their Inequalities
31
known inequalities are particular cases of this famous result. It has been the object of much research full details of which can be found in the various references;
[AI pp.10-14; CE p.953; DI pp.139-141; EMS pp.234-235; HLP pp. 70-75; PPT pp.43-57; RV p.89]. THEOREM
n
> 2,
12 [JENSEN's INEQUALITY] If I is an interval in 1R on which f is convex, if
w a positive n-tuple, a an n-tuple with elements in I then:
1
n
f ( Wn ~ wiai
1
)
n
< Wn ~ wd(ai)·
(J)
Iff is strictly convex then ( J) is strict unless a is constant. 0
It should first be remarked that the left-hand side of (J) is well defined since
the argument of
f
is the arithmetic mean of a with weight w and so its value is
between the max a, and min a, in particular it is in I; see II 1.1(2). Following later usage we will refer tow as then-tuple of weights.
(i} This proof, by Jensen, is by induction on n and the case n == 2 is just the definition of convexity, 4.1 (1); see for instance [Hrimic 2001].
0, 0 < s < 1. Since D2(0) == D 2(1) == 0 for some so, 0 <so < 1, D~(so) == 0; then 1- sox+ soy is a mean-value point for f on [x, y] 9 . Further D~(s)
=f(y)- f(x)- (y- x)f'( 1- sx + sy)
D~(s) ==- (y- x) 2 !"( 1- sx + sy)
By our assumption and Theorem 6(b), D is strictly concave and D~ is strictly decreasing. Hence s 0 is unique, with D~ positive to the left of the mean-value point, and negative to the right. Since then D 2 is not constant we have that it is positive except at s == 0, 1, which proves ( J2), and that f is convex in the sense of 4,1 Definition 1. The case n == 3 of (J) can be written as,
f( 1- s- tx + sy + tz) < (1- s- t)f(x) + sf(y) + tf(z),
(J3)
where 0 < s +t < 1, 0 < s < 1, 1 < t < 1 with equality only if x == y == z. To prove (J3) we, by analogy with the above, consider the function,
D3(x, y, z; s, t) 9 See Footnote 1.
==
D3(s, t)
==
(1-s-t)f(x)+sf(y)+tf(z) -f( 1- s- tx+sy+tz)
Means and Their Inequalities where without loss in generality we have x 0
33
< y < z, and
< s < 1, 0 < t < 1, 0 < s + t < 1.
T
Since D 3 is continuous it attains both its maximum and its minimum on T and if this occurs in the interior of T then it occurs at a turning point. Now
:s D3(s, t)
=
f(y) - f(x) -
!' ( 1 - s- t x + sy + tz) (y- x),
a
8tD3(s, t) ==f(z)- f(x)- !'( 1- s- tx + sy + tz)(z- x). So for a turning point at (s, t) we must have that
!'(1- s -tx+sy+tz)
f(x) == f(z)- f(x). y-x z-x
= f(y)-
By a 4.1 Lemma 2, (f(y)- f(x))/(y- x) < (f(z)- f(x))/(z- x). So D3 has no
turning points in T and attains its minimum on the boundary of T. However on a side of T the problem reduces to the previous case; for instance when t
== 0, D3(x, y, z; s, 0) == D2(x, y; s). Hence D3 attains its minimum value of
zero at the corners ofT, the points (0, 0), (0, 1), (1, 0), which proves (J 3 ). Now it is clear that this argument easily extends to the general case. REMARK
(i)
D
Of course D2(s) is defined for all s E JR for which (1- s)x + sy E I
and the argument given above shows that
D~
is strictly decreasing for all such s.
This implies that D 2 is negative outside [0, 1]; that is
== 0 if s == 0 1· D2 (s) { > 0 if 0 < s ' 0, 1 < i,j < n, and Laij
n
=
i=1
n
n
Laij == Lflij == Lflij == 1, j=1
j=1
i=1
and for a given n-tuple a with elements in I we define
Then F is convex on [0, 1] and if w is a positive n-tuple, t_ an n-tuple with elements in [0, 1] 1
1
n
n
1
n
1
n
f(-Laj) 3 real weights are possible as the following extension
of (J) due to Steffensen shows; see [DI p.142], [Steffensen, 1919]. THEOREM
20
[JENSEN-STEFFENSEN INEQUALITY]
If I is an interval in lR on which
f
is convex and if a is a monotonic n-tuple, n > 2, with elements in I and if w is a real n-tuple satisfying
Wn
-=1-
0, and 0 < ::: < 1, 1 < i < n,
(9)
Chapter I
38
tben ( J) bolds; further iff is strictly convex (J) is strict unless a is essentially constant.
REMARK
(i)
Condition (9) is equivalent to
rrn
TXT
_j_
1
0 , an d 0 < - wi < _ WnWn _ 1, 1 < _
In addition there is no loss in generality is assuming Wn
.< _ n.
~
> 0 when (9) is equivalent
to
0
In this proof we will assume, without loss in generality by Remark (i), that
. . . a Is Increasing.
As some of the weights may now be zero or negative it is not as obvious as in 4.2 n
Theorem 12 that the left-hand side of (J) is well defined. However
an+
n-1
W·
i=l
n
-d;.,
L: wiai
=
n i=I n
L: w:' Llai and so by (9), and the hypothesis on a, a1 < W:1 L: Wiai 0. Further if w > 0 the result reduces to (J) with a value of n the k
number of non-zero elements in w; so assume there is a k, 1 < k < n, such that Wi
> 0, 1 < i < k, and
wk
< 0. Then: (10) by(J),
Now the coefficients
Wk, Wk+l, ... , Wn
satisfy condition (9) and so we can apply
39
Means and Their Inequalities the induction hypothesis to the last two terms of ( 11) to get
Now the three coefficients Wk-1/Wn, -Wk_ 1fWn, 1 also satisfy (9) and the result follows by the case n = 3. It remains to consider the case n = 3. In this case we can assume without loss in generality that W3 > 0, w1, w3 > 0, w2 < 0 and a1 < a2 < a3; all other cases either contradict (9) or reduce to (J) in the case n = 2. Consider then (10) and (11) with k = 2; there is no need to apply (J) so they are W2a2 + w3a3 _ . , we get that W equal. Then, putting a = 3 3
_1
L wd(ai) = - :!!.!_ (f(a2)- f(ai)) + W2f(a2) + waf(ag) w3
w3
w3.
~=1
>- ~ (f(a2)- f(a1)) + f(a), 1
Hence, putting a = W 3
1 W
by(J).
3
L
wiai
i=1
3
L wd(ai)-f(a) > - ~ (f(a2) -
f(a1))
3
3 .
~=1
+ f(a)
- f(a)
== w1(a2- a1) [f(a~ -· ~(a) _ f(a2)- j(a1)] a2- a1 a- a by 4.1 Lemma 2, since a 2 < a < a3, and a 1 < a < a3.
w3
> 0,
This completes the proof of the case n = 3. ( ii) Proof (viii) of 4.2 Theorem 12, gives a particularly simple geometrical induction of the Jensen-Steffensen inequality. While D 2 ( s), using the notation of that proof, is positive precisely on the interval ]0, 1[, see 4.2 Remark (i), the function D 3 (s, t) is positive on a region larger than the interior of the triangle T. This is because D 3 is continuous and positive on the whole of T except for the corners. Precisely D3
> 0 in the region bounded by the 0-level curve of D 3 that passes
through the corners of T. The region bounded by the 0-level curve depends, in
Chapter I
40
general, on the values of x, y, z. Thus if x == y it is the strip 0 y ==zit is the strip 0
< t < 1, while if
< s + t < 1.
The question to be taken up is to find, if possible, a region larger than T that does not depend on x, y, z, as T does not so depend, and on which D 3 > 0. The region we are looking for isS== n{(x,y,z);x 1;
follows from the observations made about D2, 4.2 Remarks (i), that D3 the extensions of these sides; that is D3 (s, t) < 0 if (i) t == 0 and s
(ii)
s
== 0 and t
< 0 or t > 1 (iii) s + t == 1 and t < 0 or t > 1.
In addition considerations of the partial derivatives of D3 at the corners of T show that D 3
< 0 in the regions containing the external angles of T; that is the regions
bounded by two rays on which we have just seen that Ds is negative. The tangent to the 0-level curve at the origin makes an angle fJ1 with the positive
.
8Ds/8s (0, 0)
(y- x) (f'(c)- f'(x)) (z _ x) (f'(d) _ f'(x)), here as below
s-axis where tanfJ 1 == - aD jat (O, O) == 3 x < c < y, x < d < z; as a result we have that -1 the line s
+t
< tan fJ1 < 0. This implies that
== 0 crosses the 0-level curve at the origin, being on the side of T
when s < 0. Similarly the tangent to the 0-level curve at the (0, 1) makes an angle fJ2 with the
(y- x) (f'(c)- f'(z)) positive s-axis where tanfJ2 ==- (z _ x) (f'(d) _ f'(z)), and so -1 < tanfJ2 < 0. ..
.
This implies that the line s
+t
== 0 crosses the 0-level curve at this point, being
above the curve when s < 0. A similar discussion at the point ( 1, 0) leads to an angle with a tangent that is sometimes positive and sometimes negative depending on the values of x, y, z, since
(y- x) (f'(c)- f'(y)) there we have, with the obVIous notation that tan03 = - (z _ x) (!'(d)_ f'(y)), .
.
.
and so as is to be expected this corner is of no interest to us. At the first two corners we have locally that D3 is positive on the sides of P. We now show that in fact D3 is positive on these two sides of P.
41
Means and Their Inequalities Put ¢(s) == Dg(s, 1) when ¢'(s) == f(y)- f(x)- f'(-sx
+ sy + z)(y- x)
and
¢"(s) == -f"(-sx + sy + z)(y- x) 2 . So¢ is concave, zero at s == 0, with a unique maximum at so < 0, where -sox+ soy+ z is a mean-value point off on [x, y]. Now consider 1(s) == Dg(s, -s). A similar argument shows that 1 is concave, zero at s == 0 with a unique maximum at s1 < 0 where x
+ s1y- s1z is a
mean-value
+ h))-(f(y) -
f(y +h)),
point off on [y, z] So finally consider Dg(-1, 1) == f(x)-f(y)+ f(z)-f(x-y+z) == (f(x) - f(x
where h == z- y. Hence by the convexity off, we get that D 3 (-1, 1)
> 0.
So D 3 is positive on the sides of P except at the corners of T and so by the general properties of D3 we have that D3
> 0 on P, being zero only at the corners of T ;
in addition D 3 attains its maximum value on one of the sides of P. In particular we have, on rewriting the inequalities that define P, that if x
D2(x, y; s), 0 < s < 1; in particular the maximum value of D 2(x', y'; s) is larger than that of D 2(x, y; s). Hence the maximum of D3 occurs at (0, to) where 1 -to x + toz is the mean-value point for f on [x, z], since we have taken x == min{ x, y, z }, z == max{ x, y, z }; that is if 0 < s < 1, 1 < t < 1, 0 < s ( 1 - s - t) f (x)
+ t < 1 then
+ sf (y) + t f (z) -
+ sy + tz) 0 then f is log-convex if and only if g(x) == a
E
eax
f(x) is convex for all
JR.
D
(a) This is an immediate consequence of 4.1 Theorem 4(f).
(b) See [AI p.1 9].
(ii)
ExAMPLE REMARK
4.5.3
D
The converse of (a) is false as f(x)
== x 312 , x > 0, shows.
See also [DI p.163; RV pp.18-19], [Artin pp. 7-14].
(i)
A FUNCTION CONVEX WITH RESPECT TO ANOTHER FUNCTION
If I C
33
DEFINITION
~
is an interval and f, g : I
~
lR are continuous, and g
strictly monotonic, J == g[I], also an interval, then f is said to be (strictly) convex with respect to g if any of the following equivalent statements hold: (a) f
o g- 1
is (strictly) convex;
(b) for some function k (strictly) convex on J, f == k
== g(t),
(c) the curve x REMARK
(i)
REMARK
(ii)
y
== f(t), t
E
o
g;
I, is (strictly) convex.
Note that the curve in (c) is actually a graph; see 4.2 Remark (ii). This idea is a convenient way of expressing a property that occurs
naturally; see [HLP p. 75], [Cargo 1965; Mikusinski]. (i)
ExAMPLE
respect to
f is log-convex if and only if log is convex with or in the terminology of the 4.5.2 Definition 31 we could say f is
In this terminology
f- 1 ;
convex with respect tog as g- 1 is !-convex. ExAMPLE
(ii)
f is convex if and only if it is convex with respect to a non-constant
affine function; see 4.1 Remark (viii).
f is convex with respect tog on I if and only if for all Xi, 1 < i < 3, three points of I with x1 < x2 < x3, EXAMPLE
(iii)
1 1 1
g(x1) g(x2) g(x3)
j(x1) j(x2) > 0.
J(x3)
This is a natural extension of 4.1 (2**). Although the conditions for
f to be convex with respect to g follow from conditions
for convexity the following condition has been obtained by Mikusinski, and in a different way by Cargo.
50
Chapter I
THEOREM
34 If g E
C2 (J), and f
E
C2 (I), where I, J are as in Definition 33, and
iff', g' are never zero then a sufficient condition for f to be convex with respect tog is that
g" f" - 1, is said to be convex if a E U and b E U implies that (1 --\)a+ ,\bE U for all A, 0 < ,\ < 1. (b) If U C 1Rn, n > 1, then the convex hull of A is the smallest convex set that DEFINITION
35 (a) A set U C Rn, n
contains A.
EXAMPLE
(i)
set { x; x == points of U.
L::
EXAMPLE
(ii)
ExAMPLE
(iii)
E
If U is a finite set U
== {(x, y); x
1
wiai,
w > 0, Wk
== {a 1 , ... ak}, say, then the convex hull is the == 1}, the set of all convex combinations of the
In 1R a set is convex if and only if it is an interval. If E
f : I 1---7 1R then f is convex if and only if the set E C 1R2 , I, y > f(x)} is convex. This set is called the epigraph off,
and the definition easily extends to real-valued functions defined on subsets of 1Rn, n > 2. By replacing the interval I in 1R by a convex set U in 1Rn, n
> 2, and the points
x, y of I by points u, v of U we can easily extend 4.1 Definition 1 to functions f: U-+ JR, and 4.5.2 Definition 31 to functions f: U-+ JR+.; [DI p.63]. Jensen's
inequality, 4.2 Theorem 12, and Theorem 18 extend to this situation with the same proofs, and 4.1 Theorem 4(a), (b) is valid in the following form; see [RV pp.93, 116-117]; and 4.5.1 Theorem 30 has a natural extension.
51
Means and Their Inequalities THEOREM
36 Iff : U
~----*
JR is convex on the open convex set U in JRn, n > 2, then
f is Lipschitz on every compact subset of U and bas first order partial derivatives almost everywhere that are continuous on the sets where they exist. There is no single criterion for investigating the convexity of functions of several It is therefore useful to define some ideas that will help in special
variables.
situations, and to list several criteria for convexity. DEFINITION
37
(a) A set U in JRn is called
a
cone if u E U implies that for all
A > 0, we have A.u E U. (b) A function f defined on a cone is homogeneous of degree a if for all A > 0 we have f(A.u) == A_ a f( u). (c) If f is twice differentiable the matrix H
==
(ffj) 1 2.
A function f is convex
on U if and only if the Hessian matrix H of f is negative semi-definite; if H is positive definite on U then
f is strictly convex.
Proofs of the Theorems 38-40 can be found in the standard references; [RV pp. 94-95, 98, 103, 117], [Rockafellar pp.23-21, 30, 32-40, 213-214], [Soloviov]. REMARK
(iii)
It is easy to state an analogous result for concave, strictly concave,
functions. REMARK
(iv)
If n
==
2 then H
==
(iff
!12
j~~)
and it is non-negative definite if
and only if
!{'1 > 0,
[or!~~
> 0], and !{'1 ~~~
-
!{~2
> 0;
(21)
His positive definite if the inequalities (21) are strict; see [HLP pp.80-81]. The following examples are important for later applications.
== 2::7 1 wiaf is strictly convex on (JR+)n. This follows from the strict convexity of f(x) == xP, x > 0, p > 1, and EXAMPLE
Let w E (JR+)n and p > 1 then
(vi)
¢(a)
Theorem 40 (b) . ExAMPLE
(2::7
If w and pare as in then previous example then 7(a) == ¢ 11P(a) ==
(vii)
wiaf) 1
1
/P
is strictly convex on (JR.+)n. This follows from Example (vi) and
Theorem 40 (e) since¢ is homogeneous of degree p ..
53
Means and Their Inequalities
a~wi/Wn) is concave on (JR+)n, strictly if n > 1. The concavity is trivial if n == 1 so suppose that we have concavity for n - 1 and write (viii)
ExAMPLE
If w is as in then previous examples then x(a) == IT~
1
(22)
Now g(x) == x(wn!Wn), x > 0 is strictly concave, by the induction hypothesis
[I~ { a~wi/Wn-d is concave, and h(x) == xWn-l/Wn is strictly concave. By Theorem 40 (d) the second term on the right-hand side of (22) is strictly concave. So using Theorem 40(f) xis strictly concave on A; see also [Rockafellar pp. 21-28]. A further example of a concave function is p(a 11m), where pis a
(ix)
ExAMPLE
polynomial that is homogeneous of degree m; see III 2.1 Remark (iii). Inequality (20) can be extended in the case of homogeneous convex functions to the following inequality that is called the support inequality; [Soloviov]. THEOREM
41
[SuPPORT INEQUALITY]
Iff is
a
convex homogeneous function of the
cone U then for all u, v E U, (23) iff is strictly convex then there is equality in (23) if and only if u rv v. D
First note that f(v
+ A(u- v))
>f(v +Au)- j(Av),
by convexity,
==f(v +Au)- Aj(v),
by homogeneity.
So I (
.
! + v, u
-
) -
v -
> -
1.
lm
A----*0+
f (Jl + A(Y:. - Jl)) - f (12) \ ' A
lim f (Jl + AY:.) - Aj (Jl) A----*0+ A
f (Jl)
==f~(v; u)- f(v).
Then (23) follows from (20). The case of equality follows from that in Theorem 39. REMARK
(v)
D
Iff is concave then (rv23) holds. Further iff is differentiable then
(23) is just (24)
54
Chapter I
4. 7 HIGHER ORDER CONVEXITY
Let ai, 0
points from the real interval I then if
f :I
(a)
1 then
_an_a_o [a; f ]n n D
55
==
[ao; 1 f ] n-1
[
-
1 an; f ] n-1
Simple calculations prove these identities.
REMARK
(iii)
(26) D
Iff and g have second order derivatives then on taking limits the
identity in (a) reduces to (fg)" == gf 1 + 2! 1g1 + fg". The definition of [ao, ... , an; f] can be extended to allow for certain of the points to coalesce, the so-called confluent divided difference, provided we assume sufficient differentiability properties for
f
so as to allow the various limits to exist. If ni + 1
of the elements of then-tuple a are equal to ai, 0
1
[a; !] n
ExAMPLE
(iii)
==
< i < m,
then
ana+·. ·nrn
no '. ... nm.' ano ao . . . anrn am [ao ' . . . ' am ; f] m .
For instance it can easily be checked that
b b·f] [a,a,''
==
1
f (b)+f'(a)-2[a,b;f]
(b-a)2
.
See [Milne- Thomson pp.12-19]; [Horwitz 1995]. DEFINITION 43
n
If I is a real interval then
f :I
~----+
1R is said to be n-convex on I,
> 0, if for all choices of (n + 1) distinct points from I. [a; f]n
> 0.
(27)
f is n-concave on I. Further if (27), respectively (rv27), is always strict we say that f is strictly n-convex, respectively strictly n-
If instead (rv27) holds we say that concave. REMARK
(iv)
If n == 2 then, by Remark (ii), Definition 42 is equivalent to 4.1
Definition 1 so 2-convex functions are just convex functions. Also from Example (i) 1-convex functions are just increasing functions; and, since [a; f]o == f(ao), 0-convex functions are just non-negative functions.
Chapter I
56 THEOREM
44 (a) A function is both n-convex and n-concave if and only if it is a
polynomial of degree at most (n- 1). (b) A function f is n-convex, n > 2 if and only if j(n- 2) exists and is convex. 0
A proof of this theorem together with many other details can be found in
[Aumann €3 Haupt pp.271-291].
0
In particular iff is n-convex, n > 2, and if 0 < k < n- 2 then 1 1 1 is (n- k)-convex; ) exist, are increasing and f~n- ) < J!;- ). (v)
REMARK
fin-
f(k)
> 0 on the interval I then f is n-convex on I, and if every subinterval of I contains a point where j(n) > 0 then f if strictly n-convex on I. (b) Iff is n-convex, n > 2, and if a, bare n-tuples with distinct elements and such LEMMA
that a
45 (a) If j(n)
< b then [a; f]n-1 < [b; f]n-1;
f is strictly n-convex this inequality is strict. (c) Iff is n-convex, n > 2, it is strictly n-convex unless on some sub-interval f is
when
a polynomial of degree at most (n - 1). (d) Iff is a polynomial of degree at least n and if
j(n)
> 0, then f is strictly
n-convex. (e) If for all h -=J 0 small enough and all x, a function
f that [x, X+ h, ... , X+ nh; f]n =
then
< x < b, we have for a bounded
f is n-convex
n!~n ~(-l)n-i (:) f(x + ih) > 0,
on ]a, b[. If the inequality is always strict then
f is strictly
n-convex on ]a, b[. D
(a) This follows from Theorem 44(b) and the case n
==
2, 4.1 Theorem 6(b);
[Bullen 1971a]. (b) This is a consequence of (26); see [Bullen 1971a]. (c) and (d) follow from (b) and Theorem 44(a); [Bullen & Mukhopadhyay p.314]. (e) Since f is bounded then it is continuous, [RV p.239], and for the rest see [Popoviciu pp.48-49]. EXAMPLE
(iv)
D
Using Lemma 45(a) we easily see that ex is strictly n-convex for
all n, and log x is strictly n-convex if n is odd, and strictly n-concave if n is even,
> 0, is strictly n-convex if sign (II~ 01 (r- i)) == 1. In particular this function is 3-convex if r > 2 or 0 < r < 1, and is strictly 3-concave if 1 < r < 2 orr < 0. ExAMPLE
(v)
Similarly xr, x
Means and Their Inequalities REMARK
(vi)
REMARK
(vii)
57
The property in (b) generalizes 4.1 (3). The property in (e) generalizes J-convexity, 4.5.1, and a function
with this property is said to ben-convex (J), or weakly n-convex. Part (e) extends 4.5.1 Theorem 30. (viii)
REMARK
An extensive study of higher order convexity can be found in
[Popoviciu]; see also [DI pp.190-191; PPT pp.15-17; RV pp.237-240].
4.8
SCHUR CONVEXITY
46 A function f : In ~ JR, where I is an open interval in JR+, is said
DEFINITION
to be Schur convex on In if for all a, b E In, b--
0 for all a E In D
Chapter I
58 Iff : In
49
THEOREM
~-----+
JR, where I is an open interval in JR+, is symmetric and
convex then it is Schur convex.
D
By Remark (ii) we can assume that n == 2 when for some A, 0
b1 == (1- A)a 1
< A< 1
+ Aa 2, b2 == Aal + (1- A)a2; see for instance 3.3 Lemma 13. So
+ Aa2, Aal + (1- A)a2) == f ((1 - A) (a1, a2) + A( a2, a1)) < (1 - A)j(al, a2) + Aj(a2, a1), by convexity.
f(b) ==!((1- A)al
==f(a), by symmetry. 0
There is an extensive literature on this subject; see in particular [AI pp.167-168; BB p 32; DI pp.228-229; EM6 p. 15; MO pp.54-82; PPT pp.332-336].
4. 9 MATRIX CONVEXITY
If I is a interval in 1R then a function
f :I
~-----+
1R is said
to be a increasing matrix function of order n if for all A, B E H:/; with eigenvalues in I and A< B we have that f(A)
< f(B) 13 .
The usages decreasing matrix function and monotone function are easily defined and if a function is monotone for all orders it is said to be operator (i)
REMARK
monotone. EXAMPLE
(i)
If f(x) == xP, 0
< p < 1 or g(x) == logx then f and g are operator
increasing on IR; see [DI p.182; MO pp.463-461; RV pp.259-262]. REMARK
(ii)
The above definition can be made without the restriction to positive
definite matrices but the examples given above need positive semi-definiteness for f, and positive definiteness for g.
If I is a interval in IR then a function f : I
~-----+
IR is said to be a a convex matrix
function, of order n, if for all A, B E H:/; with eigenvalues in I and all A, 0
f(l- AA + AB) < 1- Aj(A) + Aj(B).
< A < 1, (29)
If the opposite inequality holds then the function is said to be a concave matrix function, of order n. A function that is convex, concave, of all
REMARK
(iii)
orders is said to be operator convex, concave; [DI p.64; MO pp.461-414; RV pp.259-262]. 13
The various matrix concepts are defined in Notations 7.
Means and Their Inequalities ExAMPLE
then
f
(ii)
If f(x)
59
== x 2 , x- 1 , 1/ yiX then f is operator convex ; if f(x) == yiX
is operator concave.
REMARK
(iv)
The comments in Remark (ii) apply here as well.
The topic of operator monotone and operator convex functions has a large literature; in addition to the above references see [Furuta], [Ando 1979; Kubo &
Ando].
II
THE ARITHMETIC, GEOMETRIC AND HARMONIC MEANS This chapter is devoted to the properties and inequalities of the classical arithmetic, geometric and harmonic means. In particular the basic inequality between these means, the Geometric MeanArithmetic Mean Inequality, is discussed at length with many proofs being given. Various refinements of this basis inequality are then considered; in particular the Rado-Popoviciu type inequalities and the Nanjundiah inequalities. Converse inequalities are discussed as well as Cebisev's inequality. Some simple properties of the logarithmic and identric means are obtained.
1 Definitions and Simple Properties 1.1 THE ARITHMETIC MEAN DEFINITION
.
1 If a== ( a 1 , ... , an) is a positive n-tuple then the arithmetic mean of
a 1s
(1) This mean 1 is the simplest mean and by far the most common; in fact for a nonmathematician this is probably the only concept for averaging a set of numbers. The arithmetic mean of two numbers a and b, (a+ b)/2, was known and used by the Babylonians in 7000 B.C., [Wassell], and occurs in several contexts in the works of the Pythagorean school, sixth-fifth century B.C. For instance in the idea of the arithmetic proportion 2 of two numbers, a, b, 0 that x- a : b- x :: 1 : 1, when of course x
< a < b, the number
x such
== (a+ b)/2. 3 Aristotle, also in
the sixth century B.C., used the arithmetic mean but did not give it this name. Another interpretation arises from the picturing of addition as the abutting of two line segments. One then asks what line segment when abutted to itself will 1
More precisely called the arithmetic mean with equal weights; see Remark(x) below.
2
The notion of a proportion, between four positive numbers, lengths, areas etc. is somewhat archaic but has a long history, see[Euclid Book V, Heath vol 1.pp.84-90, 384-391]. If A,B,C,D are positive numbers then A:B::C:D, read A is to Bas Cis to D, is equivalent to A/ B=C/ D. 3 In this form the arithmetic mean is one of the ten neo-Pythagorean means; see VI 2.1.4 and 1.2 below.
60
61
Means and Their Inequalities
produce the same length as the abutting of two given line segments.The idea of arithmetic mean is also found in the concept of centroid used by Heron, and earlier by Archimedes in the third century B.C; [Heath vol.II p.350]. In the notation introduced in Definition 1 the n-tuple may be replaced by some specific formulation such as 2ln (a1, ... , an), or 2ln (ai, 1
< i < n). Further if n
== 2
the suffix is omitted unless it is needed for clarity4 ; thus we will write 2l( a; w) etc., when n == 2. When there is no ambiguity either a, or the subscript n, or both may be omitted. CONVENTION
1
If a is an n-tuple, n
> 2, and if 1 < m < n, we
will write, whenever there is no ambiguity,
2tm(a) == a1
+···+am. m
CONVENTION
2
The various not at ions introduced above w i 11
apply in various contexts throughout this work. Clearly (1) does not depend on a being positive, and many of the properties of 2l can be deduced without this assumption, see below Remarks (v), (x), 1.3.3, 1.3.4, 1.3.9 and 2.4.5 Remark (v). In fact the whole discussion can take place in a very general context, see for instance VI 5 and [Anderson & Thapp]. However we have the following convention that will hold throughout the rest of this book unless otherwise indicated. A 11 n- t up 1e s and sequences w i 11 b e p o sit i v e u n l e s s s p e c i fi c a 11 y s t a t e d o t h e r w i s e ; t h a t i s i f a i s a n nCONVENTION
3
t u p 1e t h e n u n 1e s s o t h e r w i s e s t a t e d , a E (~+) n. The question of more general n-tuples will be discussed in various places later; for a situation where the elements of a are complex see [Maj6 Torrent]. 5 If the elements a are in Q then 2ln (a) E Q, and because the above definition is so elementary there is a certain point in using only the simplest tools to deduce properties of this mean. Of course non-elementary proofs besides having an interest in themselves suggest methods of generalization, and are often simpler than the more elementary approaches. Some elementary properties of 2l are listed in the following theorem; they are all obtained by easy applications of simple properties of positive real numbers. 4
There are many means that are only defined when n=2, see VI 2, and dropping the suffix makes comparisons easier to read. However the usage with n=2, or even n=l, is useful from time to time in inductive arguments. 5 Convention 3 will not apply in Chapter IV.
Chapter II
62 THEOREM
(Ad)
2 If h is a real n-tuple, a, a+ h, b n-tuples, and if .A
> 0 tben:
[ADDITIVITYl
(As )m [m-AssocrATIVITY]
(Co)
[CoNTINUITY]
lim 2tn (a
h ~o
(Ho)
(Jn)
+ h) == 2tn (a) ;
[HOMOGENEITY]
[INTERNALITY]
mina
< 2tn(a) < maxa,
(2)
witb equality if and only if a is constant;
(Mo)
[MONOTONICITY]
witb equality if and only if a (Re)
[REFLEXIVITY]
(Sy)
[SYMMETRY]
REMARK
(i)
== b;
If a is constant, ai ==a, 1
< i < n,
tben 2tn(a) ==a;
2tn (a1, ... , an) is not changed if tbe elements of a are permuted .
When m-associativity holds for all m, 1
< m < n, as it does for the
arithmetic mean, we call it the property of substitution. REMARK
(ii)
Clearly it is inequality (2) that justifies the name of mean, see [B2
p.230]. While (2) will be called the property of internality the full property given
in (In) will be called strict in tern ali ty. REMARK
(iii)
The word monotonic used in (Mo) is sometimes replaced by mono-
tone, or by isotone, see [B2 p.230]. The concept can also be considered in a strict
form, strictly monotonic, etc. REMARK
(iv)
The properties listed in Theorem 2 are not independent. For in-
stance (In) is implied by (Mo) and (Re). REMARK ( v)
It is useful to note that these properties of the arithmetic means
hold if then-tuple a is allowed to be real. Almost all the means we will discuss in this book satisfy some or all of these conditions. The properties (Co), (Re) and (In) are so basic that we would consider
Means and Their Inequalities
63
them essential in any possible definition of a mean; the property (Ho) is a very common property but does not always hold; for examples see IV. Others, such as (Ad) are in some sense characteristic of the arithmetic mean; see VI 6. REMARK
While (Ad) gives a simple relation between 2Ln(a), 2Ln(b) and 2Ln(a+
(vi)
b) it is much more difficult to obtain one between 2Ln(a), 2Ln(b) and 2Ln(a b), but ......
one is given later, see 5.3, and is known as Cebisev's inequality. A very natural extension of Definition 1 is suggested when some of the elements of
a occur more than once or, from a practical point of view, if some are considered more important than others. DEFINITION
3
Given two n-tuples a, w, the weighted arithmetic mean of a with
weight, or weights, w is
L .. n
ot
•
(
) _
~na,w-
W1a1 + · · · + Wnan __1_ w 2 a2 • -W . n 2=1 WI+···+ Wn
(3)
It is easily checked that this more general mean has all the properties listed in Theorem 2 except (Sy). Instead this more general mean has the following property: ( Sy*)[ALMOST SYMMETRY]
2Ln(a,w) is not changed if the a and ware permuted
simultaneously. REMARK
(vii)
The name arithmetic mean will normally refer to (3), and if we wish
to specify (1) it will be referred to as the arithmetic mean with equal weights. REMARK
(viii)
Using this notation Jensen's inequality, I 4.2 (J), can be written
as:
(4) A refinement and strengthening of inequality in (2) is given in the following result LEMMA
4 .(a) If a, b and w are n-tuples, then
(b) IfWn
==
1 and n
>
2, then
.
max a- 2Ln(a·, w) > minw ---n+1 -
L
(Vai - A) 2.
l 0 as was to be shown.
and
W'n == D
Result (a) is due to Cauchy, see [AI p.204], and analogous in-
equalities can be given for the geometric and harmonic means defined in the next section. REMARK
(x)
Part (b) is due to Alzer and can be extended to real n-tuples a;
[Alzer 1997d]. REMARK
(xi)
Another generalization is given in [Bromwich pp.242, 418-420, 473-
474], [Bromwicb]. REMARK
(xii)
For an amusing discussion of weighted arithmetic means see [Falk
& Bar-Hillel]. 1.2 THE GEOMETRIC AND HARMONIC MEANS
Two other means of a very ele-
mentary nature have been in use for a long time. Like the arithmetic mean they arise naturally in many simple algebraic and geometric problems, some of which are to be found in Euclid and in the work of the Pythagorean school. Thus in that . . 2ab a+ b : b; furschool's theory of harmony study 1s made of the proportions a: a+ b:: 2 ther in Euclid we have the study of the proportions x-a: b-x:: a: x when x == v-;;:6, the geometric mean of a and b, and x- a: b- x ::a: b when x
==
2abj(a +b), the
harmonic mean; see also 1.1 Footnote 3. The name geometric is probably based on the Greek picturing of multiplication of two numbers as the area of a rectangle. One then asks what number multiplied by itself will produce a square of the same area as the rectangle obtained by multiplying two given numbers; [Aumann
1935b].
Means and Their Inequalities
65
5 Given two n-tuples a, w, the weighted geometric mean, respectively
DEFINITION
harmonic mean, of a with weight, or weights, w is n
ll5n (a; w) = (
IT af'i)
1/Wn
(5)
,
i=l
respectively
(6)
REMARK
(i)
The other variations of notation introduced for the arithmetic mean
will be used here, and with other means introduced later. Thus if w is constant we get the means with equal weights, written 0
Since (2) is homogeneous 9 there is no loss in generality in assuming ab
== 1,
== 1/a when Lemma 3 is equivalent to: a+ ~ > 2 with a equality if and only if a == 1, which is just I 2.2 (20). or equivalently that b
( iv)
A variant of the the previous proof is to note that if ab
== 1, 0
0; expanding this last inequality leads to
a+ b > 1 + ab == 2. ( v)
The first geometric proof is in [Heath vol.II pp.363-364; Pappus Book 3, p.51]
and is illustrated by Figure 3.
F
Figure 3
The angles ADC, DEC and DOF are right angles
DB=b
AD=a
A 9
D
DO=l/2(b-a)
0
B
An inequality P>Q is said to be homogeneous when the function P-Q is homogeneous.
74
Chapter II
Take any point D on the diameter AB of a semi-circle of centre 0, let AD == a, DB== b. Construct the right angle ADC, then CD== v'OJj, and CO== (a+b)/2. The shortest distance from C to AB is the perpendicular distance so (}D < CO with equality if and only if D == 0. If DE is perpendicular to CO it is not difficult to check that CE == 2ab/(a +b) == 5) (a, b). So using a similar argument to the above some have proved that if a =1- b then JJ(a, b) < QJ(a, b); see [Ercolano 1972, 1973; Gallant; Garfunkel & Plotkin; Grattan-Guinness; Schild; Sullivan].
(vi)
A similar proof can be found in [Ercolano 1972] but using Figure 4.
c T Angles CON, ONT, OTD are right angles
Figure 4
A
0
Here AD
==
b, BD
== a;
N
and then OD
B
== (a+ b)/2, ND
D
2ab/(a + b), and
TD == v'OJj.
(vii) D
Another geometric proof is given in Figure 5. C
b-a Figure .5
a
A
b
B
Let ABC D be a square of side b, and let ABFE be a rectangle of sides a and b. Then areaABFE ==area AGE+ areaABFG <area AGE+ area ABC; that is
a2
b2
ab< - -+2 2'
Means and Their Inequalities
75
which is equivalent to (2). Further equality occurs only when ABC and ABFG have the same area, that is if and only if a
==
b.
Figure 6
(viii)
The arithmetic mean of a and b is the common value of the co-ordinates of
the point where the level curve of f(x, y)
== x + y passing through the point (a, b)
== x. This simple observation, together with a similar one for the geometric mean that is obtained when f is replaced by g(x, y) == xy gives a simple meets the line y
proof of Lemma 3 base on the geometry of these curves. Assume that 0 < a < b and let OGA, PGQ, PAQ be the curves y == x, xy ==
ab, x + y == a + b, respectively, see Figure 6. Then G == (~ (a, b), ~ (a, b)) and A == (2t( a, b), 2t( a, b)). Since the function f(x) == abx- 1 , x > 0, is convex, I 4.1 Corollary 7(a), the chord PQ, that is the line x + y == a+ b, lies above the graph off, I 4.1 Remark (iii), that is above xy == ab, and so the point G is to the left of the point A; that is ~(a, b)< 2t(a,b). Alternatively consider the fact that the two curves P AQ, PGQ only meet at P and Q, and that at P the slope of the first is -1, and that of the second is -ajb. Since -a/b > -1, the curve PGQ lies to the left of P AQ between P and Q. The exponential function is strictly convex, I 4.1 Corollary 7(a), so by (J), . h equa1'1ty 1.f and on1y 1.f x == y. p utt1ng . a == exp ( x + y) < exp x + exp y , w1t 2 2 ex, b == eY and noting that the exponential function is strictly increasing completes
(ix)
this proof.
(x)
A simple proof is given in [Hiisto].
If 0 1 and cB(a, b)
2t(1,u) 1 1 = cB(l, u) = 2 ( u + u)
which is not less than 1 by I 2.2 (20).
(xi)
See 5.5 Remark (ii); [Burk 1985, 1987].
-REMARK
(ii)
D
Proofs (iii), (iv)and (ix) will be adapted to prove (GA); see 2.4.2
Chapter II
76
proof (xi), 2.4.3 proof (xxix), 2.4.2 proof (xvi). REMARK
(iii)
Proof ( ii) which is in [Eves 1980 p.14] can be elaborated to give a
further inequality; see VI 2.1.4 Theorem 22. REMARK
(iv)
Proof ( x) is an elaboration of proof (iii) and proves a little more;
not only is Ql( a, b)/ Q) (a, b) bigger than 1, but if a < b the ratio is increasing as a function of b and decreasing as a function of a. A similar proof can be given by considering the difference Ql(a, b)- 1 then (rv3) balds, with the same case of equality. 0
(i)
(a} We give eleven proofs of this result. Inequality (3) can be written as (~)"'
< 1 +a(~ - 1),
which follows from
(B), I 2.1, on putting alb== 1 + x. ( ii)
It follows from a simple case of 2.2.3 Lemma 5 below that we need only
consider a and {3 rational; the following proof using that assumption is given in
[Aiyar]. Let a== Pl(p+q), {3 == ql(p+q),p, q EN* and, assuming that 0
-+p p I
(5)
'
a form of (3) used later; see III 2.1. Inequality (5) is sometimes called Young's inequality, although it is really a very special case of that inequality, [AI pp.48-49;
MPF pp.379-389].
(vii)
This is another proof of the case of rational weights; [Brown].
Let m, n be two positive integers and 0 < e < d, then em-1 + em-2d + ... + dm-1 em_ dm ---== en-1 + en-2d + ... + dn-1 en - dn
>
mem-1 ndn-1
mem --. >ndn
On multiplying this gives: ndn(dm - em) > mem(dn - en). On rewriting this inequality we get mem+n +ndm+n > (m+n)emdn. Now in this last inequality put a== em+n, b == dm+n to get (3) in the case of rational weights.
(viii)
The restriction to rational weights in the previous proof can be removed,
see [Bullen 1997].
Means and Their Inequalities
79
Consider the distinct power functions ¢1 ( x) == xu, ¢2 (x) == xv; where u =/= v, u, v > 1 and x > 0 . Applying the mean-value theorem of differentiation, see I 2.1 Footnote 1, to both of these functions on the interval [c, d], c > 0, we get, on cancelling the common factor d - c,
ve v-1' 2 for some e 1 , e 2 , with c < e1, e2 1,
or on multiplying out vdv(du- cu) > ucu(dv- cv). We have assumed that u =/= v but it is easy to check that the last inequality remains valid when u == v. The proof now proceeds as in proof (vii) with u, v instead of m, n respectively. If the Cauchy mean-value theorem 10 is used we can take e 1 = e 2 and then we need only take u, v > 0, although given the fact that we only really use the ratios uj(u + v), vj(u + v) the initial restriction, u, v
> 1, is unimportant.
The method used in proof (iii) of (J), I 4.2 Theorem 12 , can be adapted to give a proof of (3); [Bullen 1979, 1980]. ( ix)
Changing notation by putting x = a, y = b, t = (3, 1 - t G(t) = x 1 -tyt
> A(t) = (1 - t)x + ty,
=a, (3) becomes 0
< t < 1.
(6)
Further since the equality is trivial if x = y we can, without loss in generality, assume 0 < x < y. It is easy to see that both of the functions A and G strictly increase, from the value x when t == 0 to the value y when t = 1. Simple calculations give: A'(t) == y-x, G'(t) = (logy-logx)G(t); and A"(t) == 0, 2 G"(t) = (logy -log x) G(t). Then clearly A' > 0 and G' > 0, which confirms the above remark. Further however G" > 0 so G is strictly convex. Since A is linear and A(O) = G(O), A(1) = G(1), we see that the graph of A is a chord to the graph of G; and G being strictly convex it has a graph that lies below this chord, which is just (6). An alternative method for this proof is given below in the discussion of (b). 10
The Cauchy, or extended, mean-value theorem states:if the functions j,g are continuous on [a,b]
and differentiable on ]a,b[, with g 1 never zero, then there is a point c, a f(a/b), with equality if and only if t
== ajb, that is
1 with equality if and only if t == ajb. Now putting t == 1 , ==a,~ == f3 in the last p p' inequality gives (3), and the inequality is strict unless a== b. (xi)
See also the proof of VI 5 Theorem 2 (a).
(b) We divide this part of the lemma into two cases. Case (i) : a < 0 Put a' == a!3, b' == bf3, a' == -a/ (3, (3' == 1 -a' == 1/ (3. Now apply (3) to a', b', a', (3'.
a> 1 In this case f3 < 0 so the previous argument can be easily adapted. Case (ii):
Alternatively we could use (B). The proof ( ix) of (a) gives both cases immediately. With the notation of that proof write D(t) == A(t)- G(t), t E ~. Then we have D(O) == D(1) == 0, and
D"(t) < 0. Hence D(t) > 0, 0 < t < 1, and D(t) < 0, 0 > t, ort > 1. REMARK
(i)
D
It follows from proof (i) of (a), and (b) that Lemma 4 and I 2.1
Theorem 1 are equivalent. In particular proofs ( ii)-( vi) of (a) can be used to give alternative proofs of (B). 2.2.3 THE EQUAL WEIGHT CASE SuFFICES
We now show that 2.1 Theorem 1 can
be deduced from its equal weight case. LEMMA
5
It is sufficient to prove Theorem 1 for the case of equal weights.
D Once Theorem 1 has been proved for constant w, simple arithmetic arguments lead immediately to the case of rational weights; then (G A) for general real w follows by a limit argument.
To complete the proof it must be shown that if a is not constant then (1) is strict.
Means and Their Inequalities
81
Suppose that not all of the weights are rational and write wi
> 0 and
==
ui +vi, where
Vi E no. 2.3 SOME GEOMETRICAL INTERPRETATIONS
Before turning to the proofs of
(GA) we give some particularly simple forms of that inequality that have interesting geometrical applications. LEMMA
8 Theorem 1 is equivalent to either of the following.
(a) If a is an n-tuple such that IT~ only if a is constant; (b) If a is an n-tuple such that 2::~ if and only if a is constant.
D
ai
1
1
== 1 then
ai
== 1 then
2::~
IT~
1
ai
1
ai
> n, with equality if and < (1/n) n, with equality
In both cases one implication is trivial.
Assume that (a) holds ; let a be any positive n-tuple; define b = ( ~ , ... , ~). n
Clearly 2{
n
II bi == 1, and so by (a) L bi > n. i=1
>
This by the definition of b is just
i=l
Assume that (b) holds ; let a be any positive n-tuple; define c = ( :~, ... , : ; ) . n
Clearly
L
n
Ci
== 1, and so by (b)
i=1
just
Q)
II
Ci
< (1/ n) n. This by the definition of c is
i=1
< 2l.
Thus either (a) or (b) is sufficient to imply (GA) in the case of equal weights,
D
which by Lemma 5 is sufficient to prove this lemma. REMARK
(i)
The case n
== 2 of Lemma 8(a) is just I 2.2
(20); for (b) see [Goursat].
Both results are classical and proofs can be found in many places; see for instance [Chrystal pp.52-56], [Darboux 1887].
Both parts of this lemma have simple geometric interpretations. COROLLARY
9 (a) Of all n-parallelepipeds of given volume the one with the least
perimeter is the n-cube. (b) Of all n-parallelepipeds of given perimeter the one with the greatest volume is then-cube. COROLLARY
10 Of all the partitions of the interval
[0, 1] into n sub-intervals, the
83
Means and Their Inequalities
partition into equal sub-intervals is tbe one for which the product of tbe intervals is the greatest. See [P6lya 1954 p.129]. LEMMA
11 (a) In the case of n
== 3 (GA) is equivalent to tbe statement: of all tbe
triangles of given perimeter tbe equilateral triangle bas tbe greatest area. (b) In tbe case of n == 4 (GA) for 4-tuples of numbers tbe sum of any three of wbicb is greater tban tbe fourth is equivalent to tbe statement: of all tbe concyclic quadrilaterals of given perimeter tbe square bas tbe greatest area. (a) Let a, b, c be the lengths of the sides of a triangle, s == (a+ b + c)/2 its semi-perimeter; then by a formula of Heron, [Heath vol.II pp. 321-323; Melzak 0
1983b pp.1-3], its area is A
==
J s(s- a)(s- b)(s- c).
equilateral this area is Ao == s 2 /3v'3.
When the triangle is
By (GA) in the case n ==
3 and
equal
weights,
A=
vs( {/(s- a)(s- b)(s- c))3/2 0, b + c + d- a== 2a 1 > O,c + d +a- b == 2a2, d +a+ b- c == 2as so that a, b, c, d
Then simple calculations show that 2s
are the sides of a concyclic quadrilateral, see [Melzak 1983b pp.B-9]. Now using the inequalities above Q;4(a1, a2, as, a4)
== A 112 < A~ 12 == s/2 == 2s/4 == Ql4(a1, a2, as, a4),
with equality if and only if a == b == c == d, or a1 == a2 == as == a4. Noting that Lemma 5 shows that (GA) for equal weights and a given n-tuple implies the general case for the same n- tuple completes the proof. REMARK
(ii)
D
Extending this to a general concyclic n-gon seems difficult as there
seems to be no simple formula for the area of a concyclic n-gon, n
>
5; see
[Robbins]. The need to phrase part (b) of this lemma in a manner different from part (a) was communicated to me by Mowaffaq Hajja. REMARK
(iii)
For further discussions of the results in this section the reader
should consult the following: [Kazarinoff 1961a pp.18-58; Kline p.126], [Bioche; Garver; Sbisha; Usai 1940b]. 2.4 PROOFS OF THE GEOMETRIC MEAN-ARITHMETIC MEAN INEQUALITY
The
proofs will, as far as is known, be in the order of their appearance. Seventy four proofs are given and there are undoubtedly as many more in the literature. A survey of proofs given up to 1904 can be found in [Muirhead 1901/04] and a survey of several proofs also occurs in [Colwell & Gillett]. It is sufficient by Lemma 5 to give a proof for the equal weight case and many proofs do this. Also, given 2.2.1 Lemma 3, or 2.2.2 Lemma 4, it is sufficient to give the inductive step in any proof by induction; see also 2.2.4 Remark (i). Further for a complete proof of the theorem it is clearly sufficient to prove that the inequality is strict for non-constant n-tuples. In addition in an inductive proof we can also assume that no two elements of the the n-tuple a are equal, for if two are equal the inequality follows by the induction hypothesis; in fact we may assume that the n-tuple is strictly increasing. The proofs are divided into sections determined by the publication dates of the four main references [HLP; BE; AI; MI], 1934, 1965, 1970 and 1988, respectively,
Means and Their Inequalities
85
preceded by the prehistoric proofs. A baker's dozen of proofs that are in journals not seen by the author are listed for completeness in section 2.4. 7. 2.4.1 PROOFS PUBLISHED PRIOR TO 1901. PROOFS (i)-(vii)
( i)
1729
MACLAURIN CIRCA
This is by far the earliest proof and it appears to be due to Maclaurin who states the result in the form of 2.3 Corollary 10.
D Suppose that 0 < a1 < a2 < · · · < an, a1 =f. an. If a1 and an are replaced by (a 1 + an)/2 then 2l is unchanged, but using (2) it is easily see that Q5 is increased. If then a is varied so as to keep 2l fixed, and of such n-tuples a' is the one at which Q5 assumes its maximum value, the above argument shows that a' must be constant. Hence the maximum of Q5 is attained when all the terms of a are equal,
D
and this maximum value is equal to 2l.
[Chrystal p.47], [Grebe; Maclaurin]. Given the date of the proof it is not surprising that the existence of an a' at which Q5 attains its maximum was taken for granted. This missing step can be supplied
in either of the following ways, [HLP p.19, footnote( a)]. (a) Use the fact that a continuous function, ¢,defined on a compact set, K, attains its maximum on that set. In this situation if x == (x 1 , . . . , xn), n
¢(x)
=(IT x1)
1/n
; K
={x;
n
Xi>
0,1
< i < n, and
i=l
L:>i =n2t}. i=l
(b) After k steps in Maclaurin's proof let us denote the resulting n-tuple by a(k),
and assume it is increasing. Then clearly a(k)
< a(k+l) < ... < a(k+l) < a(k). -n -n
1-1-
Hence limk~oo a~, and limk~oo a~ both exist, say with values
a,
A respectively. It
is not difficult to see that after n steps that maximum difference in the sequence has been reduced by at least one half; that is
. ( an(k) and so 1Imk~oo
( ii)
CAUCHY
-
a 1(k))
--
0 , or A -- a. This argument Is . d ue to H ar d y.
1821
This elementary proof depends on a sophisticated induction argument, see 2.2.4 Lemma 6, and consists of proving (GA) for all integers n of the form 2k, k EN*.
86
Chapter II
D
Assume the result is known for k == m and let a== ( a1, ... , a2~, b == (b1, ... , b2~
and
C
== (C1, ... , C2m+l) == (a, b) == (a1, ... , a2m, b1, ... , b2m).
Now
~2m+1 (c) ==J~2m (g_)~2m (Q),
<J2l 2m(a)2l2m(b), by the induction hypothesis
(7)
®~(a),
the second term being positive by internality.
After at most (n- 1) repetitions of this process we arrive at a constant n-tuple a 11 , and this gives a proof of (G A). D [Briggs & Bryan p.185; Hardy 1948 p.32; P6lya 1954 p.247; Sturm p.3], [Craw-
ford; Fletcher; Muirhead 1900/01, 1901/1904, 1906].
and an
A similar proof has been constructed by defining a~ == ®n(a) == ®, a 1 a2 / Q). The idea in this proof has been as a basis of a general method
(v)
REMARK
==
of forming inequalities, [Keckic].
(vii) D
CHRYSTAL
1900
Again let then-tuple a be increasing and non-constant, and assume (GA) for all integers less than n. o(
~n
(
a)
== n -
n
1 ot
~n-1
(
a)
+ an n
==
ot
~n-l
(
a)
+ an -
2ln -1 (a) · n
(9)
By internality an - 2tn_ 1 (a) > 0 and so from the right-hand side of (9)
2t~ (a) > ~~ _1 (a)
+ (an
- ~n-1 (a)) 2(~
=i (a),
by ( B), f'J
==an2l~=i (a)
>an Q)~=i (a), by the induction hypothesis, ==®~(a).
0
[Chrystal val. II], [Muirbead 1901/04; Oberscbelp; Popovic; Wigert]. REMARK
(vi)
If the inductive hypothesis is used on the middle expression in (9)
then we get ot
n (a ) > ( t1.t ( ) Vn-1 a
~n
+ an -
® n-1 (a) ) n
n
.
89
Means and Their Inequalities and the above argument again gives (GA); see [Weber pp.689-690], [Tweedie]. 2.4.2
PROOFS PUBLISHED BETWEEN
(viii)
MUIRHEAD
1901
1934.
AND
PROOFS
(viii)- (xvi)
1900/1901
See V 6 Remark (iii).
( ix)
DOUGALL
1905
Following ideas of Muirhead, Dougall made a study of various identities from which (GA) becomes apparent. In particular he gives the following proof of (GA). With the notation of I 1.1(1) and Example (i) , if a is a non-constant n-tuple
D
the polynomial
has distinct roots so inequality I 1.1 ( 4) is strict, that is
> dl/r+l dl/r n-r-1' n-r In particular dn- 1
1
< r < n.
> d~/n, which is (GA).
D
[Dougall]. REMARK
In fact Dougall gives an exact formula for d'l-dn; such a formula also
(i)
occurs in [Jolliffe; Muirhead, 1900/01]. A similar proof can be found in [Green].
(x)
P6LYA
D
1910
If a is a non-constant n-tuple define b by ai
Then b is not the zero n-tuple and L~
1
wibi
== (1 + bi)2tn(a; w), 1 < i < n.
== 0, and
n
L wibi), by inequality I 2.2(8), i=l
0 [HLP p.103], [Alexanderson], [Lidstone; Wetzel]. REMARK
(ii)
This proof is in [HLP] but is assigned this date by Alexanderson.
The equal weight case was rediscovered by Lidstone. REMARK
( xlii).
(iii)
An extension of the idea in this proof is given below in 2.4.5 proof
Chapter II
90
(xi)
DoRRIE
1921
If a is a non-constant n-tuple, n
0
> 3, and
TI~
1
ai
== 1, then at least one
element is bigger than 1 and at least one is less than 1; assume a 2 < 1 < a1. Then n
n
Lai ==a1
i=1
+ a2 + Lai i=3
n
> 1 + a1a2 +
L ai,
see 2.1 Lemma 3 proof ( iv ),
i=3
>1
+ (n- 1) == n
by the induction hypothesis, 0
and by 2.3 Lemma 8(a), this is equivalent to (GA).
[Dorrie pp.37-39; Korovkin p. 7], [Ehlers; Heymann; Kreis 1946].
(xii)
CARR
0
The following is a simple inductive proof.
1926
= ( 1!3;/n-1 (a )a~-2/n-1) 1/n-1 (!j~:._:~/n-1 (a)
IBn (a)
1 + n- 2 Q) _ (a) ®1/n-1(a)an-2/nby (5), the case n == 2, -n n- 1n n- 1 n1_, 1 n-2 n-2 < (n _ 1) 2 ~Bn(a) + (n _ 2) 2 an+ n _ 1 ®n-1 (a), by (5) again, 1 n-2 n-2 < (n _ 1)2 ~Bn(a) + (n _ 2 ) 2 an+ n _ 1 2tn-1 (a), by the induction hypothesis, 1
-
n (
i=l
IJ ai)
l/n
+ an+l,
by the induction hypothesis,
i=1
n == 1/n + an+1 an+1 >n + 1, by (rvB) with a== -1/n and 1 + x == an+l·
D
[Newman D J]. (xxviii)
DIANANDA
1960
This is a modification of Cauchy's proof, 2.4.1 proof ( ii).
99
Means and Their Inequalities D
~
v~
( ) _
. /
"-'n a -
( ) 1/n-1~n-2/n-1( ) "-'n-1 a an "-'n a
< ~ ( 1Bn~ 1 (a)+ a;_ln~ 1 \B~~ 2 jn~ 1 (a)), < ~ ( 2tn~ 1 (a) + n a:_ 1 + ~
=~
by 2.2.1 (2), (n
= 2 equal weight
( GA)),
IBn (a)) , by 2. 2. 2 ( 5), (n = 2 ( G A)) ,
and the induction hypothesis,
n
n-2 2(n- 1) 24,( a)+ 2(n- 1) 1Bn(a), D
which on simplification gives the result. [Diananda 1960].
(xxix)
KOROVKIN
1961
In Korovkin's book I 2.2 Lemma 6 is proved using (GA). We use the lemma to give a proof of (GA) in the form of 2.3 Lemma 8(a). Let a be an n-tuple with IT~
D
1
ai
== 1 and set
Then the left-hand side of I 2.2 (24) is just 2:~ L~
1
ai
1
Xk
==
IT~ k ai, 1
n as had to be proved.
D
[K orovkin p. 8].
(xxx) D
MOHR
1964
Let a be a non-constant increasing n-tuple and first we introduce some
notation: a Co)
== a; aC 1) == (aC 0 )) ( 1 ) == (a~ 1 ), ... , a~1 )), where a~ 1 ) ==
2tn_ 1 (a~),
where a~ is defined in Notations 6(v); and having defined a(o), ... a(k- 1 ) then a(k) = (
a(k~l)) (1).
In other words the terms of -a(k) are the arithmetic means of the terms of -aCk- 1) taken (n- 1) at a time. The following identities can easily be established: 2tn(a(k)) ( k) _
ai
- 24,
( ( o) )
a
+
(_
1
== 2tn(a( 0 )),
0 ) k -1 2tn (a ( ) ) -
(n _ )k
( (O)) an - al < ot ~n a + (n _ )k' 1
1
k
k
== 0, 1, ... ;
ai
.
_
, 1 < z < n, k - 1, 2, ... ,
== 1,2, ... , by interna1ity.
Again by internality,
(18)
100
Chapter II
Now assume (GA) for all integers less than n, and we have immediately that a~k) >~ ~ _ n _1 ((a(k-1))~) _ ~
==(til;. (a (k-1)))n/n-1 ( a~~k-1)) -1/(n-1) , k == 1, 2 , .... Vn
On multiplying these inequalities we obtain (19) In addition since
a1
i= an,
From (18) and (19), (20) Using (18), choose a k so that
This last inequality and (20) with m
== 1 completes the proof.
0
[Mohr 1964]. REMARK ( v)
(xxxi)
In fact this proves a little more:
BECKENBACH
& BELLMAN 1965
This is another calculus proof of 2.3 Lemma 10(b). The object is to find the minimum of the function L~
D
set {a; a > 0, IT~ 1 ai == 1}. Using the Lagrange multiplier11 approach, consider f(a) when
a1 aa. J
. 8f
So If - 8a1 11
Iff and
==
IT ai -
.< n.
==
1
ai on the compact
IT~
1
ai -A L~
1
ai,
A, 1 < J
i=1 i:f:j
== · · · ==
8f Ban
we must have ai
== · · · == an.
g, are functions of n variables, and we wish to find the extrema off subject to condition
g=O the auxiliary function f(x,:A)=f(x)+:Ag(x) of (n+1) variables is introduced and the problem reduces to finding the turning points of j. The extra variable >. is called a Lagrange multiplier and the procedure is called the method of Lagrange multipliers; [CE p.1015; EM5 p. 336].
101
Means and Their Inequalities This gives n as the unique minimum of 2::~
1
ai
and proves 2.3 Lemma 10(b). D
[BB p.5], [Amir-Moez; Rodenberg]. 2.4.4 PROOFS PUBLISHED BETWEEN 1966 AND 1970. PROOFS (xxxii)-(xxxvii)
(xxxii)
D
DZYADYK
1966
The following two identities are given by Dzyadyk:
(21) (22)
where Pn is the polynomial of I 1.2(a) .
By I 1.2(6) the identity (21) implies that
niBn~~~ an+l >
1Bn+l(a) and, by
the induction hypothesis, the left-hand side of this expression is not greater than
nm.,.~~+l an+l
m.,.+l (a).
=
This proves (GA), and for equality we must have
a 1 == · · · ==an, by the induction hypothesis, and Q;n(a) == Q)n+l(a), by I 1.2(7). This implies that
a1 == · · · == an+l·
Identity (22), again using I 1.2(7), immediately implies (GA) and the case of
D
equality.
[Dzyadyk]. (xxxiii)
GuHA 1967
Guha uses the following lemma to give a simple proof of (GA). LEMMA
14
If p > q > 0, x > y > 0 then
(px + y + a)(x + qy +a)> ((p + 1)x +a) ((q + 1)y +a), with equality if and only if x == y. D
This is an immediate consequence of the identity
(px + y + a) (x + qy + a) - ( (p + 1) x + a) ( (q + 1) y + a)
==
(px - qy) (x - y). D
102
Chapter II
D
Repeatedly using Lemma 14 gives: n factors
n-2 factors
>(2al
+ a3 · · · + an)(2a2 + a3 +···+an) (al +···+an) ...... (a1 +···+an)
> ..... . >na1(2a2
+ a3 + · · · + an)(a2 + 2a3 +···+an) ...... (a2 + · · · + an-l + 2an) r,__
n factors
'
___..~
> · · · · · · > (na 1 ) ... (nan) == n n Qj n (a) n . The case of equality follows from that of the lemma.
D
[Guha].
( XXXiv)
GAINES
1967
The following result is well-known, [DIp. 75]. 15
LEMMA
[ScHuR]
If Ai, 1 < i < n, are the eigenvalues of the complex matrix
A== (aij )lan. a;
(J
ai) 1 /n in the last inequality gives (GA). The case of equality is
immediate. D
D
As a result of Lemma 19 to prove (GA) we need to prove (32),which we do
using a short version of Teng's proof due to Hering; [Hering 1990]. The case n
==
1 is just I 2.2(20) so suppose (32) holds for n, n
> 1. Then
114
Chapter II
I17+{ ai >
by the induction hypothesis. Now either
1 when at least one ai
>
1,
IT7+11 ai
2, a== (a1, ... , an-1, x), s == a1 + a2 + · · · + an-1, p == a1a2 ... an-1 and define f(x) == nn(~~(a)- QJ~(a)),x > 0. Then f (x) == (x + s) n - n npx, f 1 ( x) == n (x + s) n- 1 - n n p, f 11 ( x) == n (n - 1) (x + s) n- 2 . Hence f is seen to have a minimum at x == xo == np 1 /(n- 1) - s, and simple calcu0
Let n
lations give that f(x 0 ) == nnp(s- (n -1)p 1 /(n- 1 )). Now the induction hypothesis is s > (n- 1)p1/(n- 1) with equality if and only if a1 == · · · == an_ 1- a property that clearly holds when n==2, 3. So by the induction hypothesis f(x 0 ) equality if and only if x
== x 0 .
>
0, with
0
[Minassian]. REMARK ( v)
Minassian points out that this argument allows for consideration of
certain real a if (GA) is written in the form ~~(a; w) >~~(a; w). ( lxii)
0
Yu 1988
In the right-hand inequality of I 2.2 (9) put x == ai/QJn(a; w), and multiply
by Wi, 1
< i < n,
and assume that Wn
This gives Wi log ai/~n(a; w)
==
1.
< wiai- wi,
1
< i < n.
Adding over i gives (GA),
and the inequality is strict unless a is constant by I 2.2 (9).
0
[Yu]. REMARK
(vi)
This proof is similar to proof (Zviii).
2.4.6 PROOFS PUBLISHED AFTER 1988. PROOFS (lxiii)-(lxxiv)
( fxiii)
0
8CHAUMBERGER
1989
Put x == aie/QJn(a), 1
< i < n,
in I 2.2 (7) and multiply to get
On simplifying this gives the equal weight case of (GA). The case of equality is immediate from that of I 2.2 (7) .
[Scbaumberger 1989].
0
Means and Their Inequalities (lxiv)
BEN-TAL, CHARNES
&
TEBOULLE
115
1989.
See VI 4.6 Remark(vii).
[Ben-Tal, Charnes & Teboulle]. (lxv)
ALZER
1991. ....
See 5.5 where an interesting proof is given using Cebisev's inequality.
[Alzer 1991a]. (lxvi)
SCHAUMBERGER
&
BENCZE
1993
Assume that we have the following inequality: if x is an increasing non-negative
n-tuple then (34)
with equality if and only if x is constant. Now, in this inequality, put a 1 == x 1 , a 2 == 2x2 - x1, a3 == 3x3 - 2x2 ... to get the equal weight case of (GA), together with the case of equality. To prove inequality (34) note that it is trivial if n for some n
>
== 1 and assume that it is valid
1; then
X1 (2x2- x1)(3x3 - 2x2) · · ·(nxn - n- 1xn-l) ( n <x~(n
+ 1xn+l -
nxn)
+ 1xn+ nxn), by the induction hypothesis, by I 1.2(7). n + 1 Xn+l - n) < x~t~,
==x~+ 1 (
1 -
Xn
The case of equality follows from that of I 1.2(7). An alternative proof of inequality (34) is given in 3.7 Example (i).
[Schaumberger & Bencze 1993]. ( lxvii)
0 1 0, then by the right-hand inequality of I 2.2 (9) we have for each i, 1 < i < n, ai
W ~· - -
X
ai
> - w·log-· ~ '
w·~
X
further this inequality is strict unless ai / x
== 1. Adding and noting that at least
one ai/x =1- 1, we get
2tn(a;w)>x+xlog
sup F(x) xEM
with equality in either inequality if and only if all the extrema occur at the same value of x. To prove (GA) assume let a and w be n-tuples with Wn
== 1.
== Wiai(x -logx), 1 < i < n,x EM== JR.+ then fi has a minimum at Xi== 1/ai, 1 < i < n, giving I:~ infxEM fi(x) == 1 + log(Q)n(a;w)). F(x) == 2tn(a;w)x -logx which has a minimum at x == 1/2tn(a;w) giving infxEM F(x) == 1 + log(2tn(a; w)). Using the lemma we get (GA) together with Let fi(x)
1
the case of equality.
[Sandor & Szabo; Pecaric & Varosonec; Usakov]. ( lxx) HRIMIC 2000 0 Let a be ann-tuple and define then-tuple b by, bt = ... , bn
=
a1 a2 ... an l!'i~(g)
1 n < b1bn
= 1.
1 + b2b1
.
.
.
a1
( ),
Q)n a
b2 =
.
a1a2 2 Q)n(a)
, •..
A s1mple apphcatwn of I 3.3 Corollary 17 g1ves,
1 + · · · + bn bn-1
from which (GA) is immediate.
D
Means and Their Inequalities
117
[Hrimic]. ( lxxi)
SCHAUMBERGER
2000
This is another proof that uses the the right-hand inequality I 2.2 (9); see also 2.4.4 proof ( xxxvii), 2.4.5 proofs ( lviii), ( lxii) and( lxviii).
D
n
>~log ((!';:(g)),
by the the right-hand inequality I 2.2 (9),
n
=log
f!
ai ) ( (!';n(Q) = 0.
D [Scbaumberger 2000]. (lxxii)
GAO
P 2001
Assume without loss in generality that Wn
D
== 1 and
a is an increasing n-
i terms
tuple with
a1
=!=-an and write xi==
== Qtn (Xi; W)
D i (X)
- Qj n (Xi; w), , 0
(~, ai+ 1 , ••• an) and define the functions
< X < ai+ 1 , 1 < i < n.
Then
Hence Di is strictly decreasing and so
with at least one of these inequalities being strict; but D 1 (a 1 )
> 0 is just (GA). D
[Gao P]. _ ( lxxiii)
D
ROO IN
2001
This proof proves the equal weight case of (GA) for rational n-tuples a. It
is then easy to see that we can assume this n-tuple consists of integers and even that Qtn(a)
== nk where k EN*, k > 2.
Now the set A == {a; a E (N*) n and ~ (a) == nk}, for a given positive integer k is finite. Hence A contains an a' for which the product of its elements is maximum; that is \1 a E A
n
n
i=l
i=l
II ai < II a~.
Chapter II
118 Suppose that
a'
is not constant then there exist two elements of
a',
without loss
in generality a 1 ' and a~, such that a~ < k < a~, which implies that a~ -a~ and a~ - a~ - 1 > 0 . Now a" == (a~
+ 1, a~ -
> 2,
1, a~, ... , a~) E A and
i=1
n
>IT a~. i=1
This is a contradiction and so a' is constant. Hence a'1 == · · · == a'n == k and this implies (GA). (ii)
REMARK
D It is not immediate how to use this result to get (GA) for all positive
n-tuples. [Rooin 2001c]. ( lxxiv)
HAS TO 2002
This proof by induction is an elaboration of 2. 2.1 Theorem 3 proof (x), in the form suggested there in Remark (vi), and is related to Liouville's proof, 2.4.1 proof (iii). Assume without loss in generality than a ann-tuple such that
D
a1
< · · · < an.
Let X== an and put f(x) == mn(a;w)- Q5n(a;w) when
f is increasing if x > Q5n_ 1 (a; w ), in particular by the internality of the geometric mean if x > an_ 1. Hence f(x) > f(an-1) == mn-1 (a~; w')- Q5n-l (a~; w') Hence
where w is then -1-tuple (w1, ... , Wn-2, Wn-1
+ wn)·
The result follows from the 0
induction hypothesis.
[Hasto]. REMARK
(iii)
As in the proof in 2.2.1 this gives a little more as it shows 2ln (a; w)-
Q5n (a; w) is increasing as a function of any term exceeding the geometric mean of the remaining terms, and is decreasing as function of any term that is less than the geometric mean of the remaining terms. A similar proof, that also gives a little more, can be based on the ratio REMARK
(iv)
Qtn (a;
w) / Q)n (a; w).
Reference should be made to a slightly different use of the function
f in proof (iii) of 3 .1 Theorem 1. 2.4.7
PROOFS PUBLISHED IN JOURNALS NOT AVAILABLE TO THE AUTHOR
These are listed in chronological order.
Means and Their Inequalities
119
[Schlomilch 1858a, 1859], see III 3.1.1 Remark (iv); [Unferdinger 1867; Tait 1867/69; Schaumberger 1971, 1973, 1975a, 1985b, 1987, 1990b, 1991, 1995; Moldenhauer; Pelczyriski 1992]. 2.5 APPLICATIONS OF THE GEOMETRIC MEAN-ARITHMETIC MEAN INEQUALITY We have already seen a simple but effective use of (GA) in Heron's method, 1.3.5 above. Another application was given by Kepler. He noted that an ellipse with semi-axes a and b has the same area as a circle with radius -JO;b. Hence from the classical isoperimetric property the perimeter of the ellipse is bigger than the Using (GA) Kepler gave 1r(a +b) as an
circumference of that circle, 21r-JO;b.
approximation for the perimeter of the ellipse 16 . We give some further uses below; of course there are many more, see for instance [Monsky; Ness 1967]. EXAMPLE
Determine all the polynomials xn+an_ 1 xn- 1 +· · ·+ao, with ai == ±1,
(i)
< i < n-1, having all real zeros. D Suppose that the zeros are xi, 1 < i < n, then I:7 1 == a;_1 - 2an-2 == 1 ± 2 == 3, since the sum must be positive; further I17 1 == a6 == 1. By (GA) 1 < 3/n, son< 3. It is now easy to look at the finite number of possible polynomials particularly since when n == 3 the equality case of D (GA) implies that the zeros are all ±1, which gives a1 == -1; [Boyd]. 0
xr
xr
ExAMPLE
(ii)
The proof of a surprising result due to Simons, [Simons], is based
on (GA). Let a be an n-tuple and a a real n-tuple all of whose terms are distinct
and define f(x) ==
L:7
aixai, x > 0. If g(x) == axa is chosen to give a good approximation to f at the point x == xo in the sense that f(xo) == g(x 0 ) and f' (x 0 ) == g' (x 0 ) then f(x) > g(x) with equality if and only if x == x 0 . 1
The conditions imply that a
D
==
L:7
1
aiaixgi / f(xo) and a == f(xo)x 0a.
Hence
-
f (X) -
I:~
1
aixgi(x/xo)ai
L.:n
.
CXi
,;_1 a~xo ~-
ITn (X )aiaixo/f(xo) j (XQ) > f (Xo), by (GA) . xo ~=1
=(:J
a f(xo)
There is equality if and only if ((xjx 0 )ai, 1 < i < only if x 2.5.1
== xo.
CALCULUS PROBLEMS
= =
n)
axcx
g(x).
is constant; that is if and D
Problems solved by elementary calculus can often be
solved using (GA), usually in its simplest forms 2.2.1(2) or 2.2.2(5); [Niven], [Boas
& Klamkin; Frame; Lim]. 16 The classical isoperimetric property is: of all convex closed curves of a given area the one with the least perimeter is the circle. It is known that the perimeter of an ellipse cannot be expressed in terms of elementary functions; [CE p.521; EMS pp.206-207].
120
Chapter II
EXAMPLE
vrn == 1; [Rooin 2001a]. Assume that n > 3, then:
limn~~
(i)
O< V'n-1== -
n-1
n-1
< , by (GA), 1 + n1/n + ... + nCn-1)/n - n vn1/n+2/n+···+(n-1)/n n-1
1
== n(3n-1)/2n < - {Iii.
Alternatively, [Sinnadurai 1961]: n2 -n terms ...---""
1
w~, where
1
k-
vv:
n
L
Wi+lo
n i=l
For further discussion the reader is referred to Anderson's paper. 2.5.3 PROVING OTHER INEQUALITIES
This is of course one of the main applications
of (GA), and much of the rest of this book is evidence of this. Here we give a few isolated examples of this use, many more can be found in [Herman, Kucera €3
Simsa pp.151-166].
(a) It is possible to compare the geometric and arithmetic means with different weights; [Bullen 1967; Dragomir & Gob 1997a; Iwamoto; Mitrinovic & Vasic 1966a,c; Wang C L 1979d].
121
Means and Their Inequalities THEOREM
21 If a, u, v are n-tuples then:
(a) Qj n (U; U) (!) ( . ) n v, U n
Vn ) ( (!5n a; u < U
2tn
(
)
a; v ,
with equality if and only if a v u- 1 is constant;
(b)
2tn(Q; ~)) Un (2tn(Q; 12_)) Vn < ( 2tn(g; 1f + ( QJn(a;u) D
Q))
Un+Vn .
QJn(a;u+v)
QJn(a;v)
(a) It is easily seen that
which implies the result by (GA). (b) This is an immediate consequence of the similar inequality associated with (J), D I 4.2 Theorem 15(b), applied to the negative log function.
(/3) The following theorem is in [Lupa§ & Mitrovic]. THEOREM
If a is an n-tuple then
22
(36)
There is equality if and only if a is constant.
By 1.2(7) and 1.2(10) it suffices to prove (36). The right hand inequality is an immediate consequence of (GA) and the observaD
tion that 2tn(e +a)== 1 + 2tn(a). For the left inequality consider, n
n
IT (1 + ai) ==1 + L L!(ai
1 ...
airn)
m=l m
i=l
>1 +
=1 +
t (:) t (:) (n
g!(ai 1
•••
aim?!(;;',), by (GA),
aj)m/n = (1
+ ~Bn(a)t.
The case of equality follows from that of (GA) .. REMARK
(i)
Another proof of (36) has been given by Keckic, [Keckic].
D
122
Chapter II
REMARK
(ii)
A part of this inequality has been generalized by Pecaric, see I 2.1
Remark(v), and similar but more general results can be found in some papers of Kovacec and others; [Alzer 1990o; Kovacec 1981a,b; Mitrovic 1973; Pecaric 1983a]. In particular it is immediate from the above proof that the right-hand inequality in (36) holds for weighted means. See also III 3.1.3 Remark (iv).
(!) The following inequality is in [Kalajdzic 1970]. THEOREM
23
Let a, w be n-tuples with Wn == 1 and let b > 1, then:
with equality if and only if a is constant.
Let b == lfo.. == ( ba 1 ,
D
••• ,
ban),
w ==
( wh, ... , Win) and consider the sum on
the right-hand side of the above inequality
n
< L'mn(b; w)
==
(n- 1)! L bai. i=l
D
The case of equality is immediate.
(5) (GA) can be used to give a simple proof of a special case of I 2.2(11); the case when p == n
+ 1, q == n;
[Forder; Georgakis; Goodman; Melzak; Sinnadurai 1961]. n terms
1/(n+l) (
1(1+X)n ) n
1 < (1+(1+x)+· .. +(1+x)), by(GA), n+1 n n X
==1
+ n+1 '
that is
REMARK
(iii)
This shows
(1 + ~ t n
increases strictly with n, and a similar ar-
gument can be used to show that (1 2.2(a).
+ -n1) n+l
.
.
str1ctly decreases w1th n; see I
(E) Here we collect various results, without proofs.
Means and Their Inequalities THEOREM
(a)
24
[MIJALKOVIC & KELLER; LYONS]
(b)
(c)
[ZACIU]
(DAYKIN & SCHMEICHEL]
If an+k == ak, bk == Ak+n - Ak, 1 < k
(d) (AI, If ~.ia
P.
< n,
then
348]
> 0, 0 < j
0; (8) 0 < lim infn~oo an < lim SUPn~oo an
n ( 2tn (a) -
Q) n (a)) == p ( { 1, ... , n - 1, n}) > p ( { 1, ... , k, n})
i=1
Letting n
-t
oo, and then letting k
-t
oo we get that limn~oo n (2tn (a) -
==
2::~ 1 ai, which completes the proof in this case.
(/3)
First remark that this hypothesis implies: lim 2tn(a) ==a,
and
n~oo
Now define
Tij
==
ai - 2tj (a); obviously
2: i=1
Now: j
2:::~= 1 Tij
0, and we investigate the
J
J
properties of the two quantities
1 n lim - "(ai- a) 2 == 0. n~oo n L.-t i=1
Ti;,
2: Ti~· i=1
j
j
2: Ti~ + j(2tj(a)- a) > 2: Ti~
2
2
l:(ai- a) == i=1 i=1 k
>
i=1 k
2: Ti~ == 2: ((ai- a)+ (a- 2tj(a)) i=1
i=1
2 ,
i
n 0 then
i=l
r~+la-~j(a)!Lr~.
(17)
E.
.
J
Lri~ < Lrlj (lai- al
+ Ia- ~j(a)l)
i=l
no
j
j
0 choose no such that if n >no then lan- al
P({PI , QI , · · · , Pi, Qi})
>p( {PI, Ql, · · · ,Pi-I, Qi-I}) + P( {Pi, Qi}) · · · · · ·
and so limn-+oo n(2tn (a) -
and hence is negative outside the interval [0, 1]. This completes the proof. REMARK
(ii)
D
A particular case of arithmetic and geometric means with general
weights that do not satisfy (49) is discussed below in 5.8. ExAMPLE
(i)
A simple use of Theorem 23 is to prove inequality 2.4.6 (34); let x
be an increasing n-tuple then
3.8 OTHER REFINEMENTS OF THE GEOMETRIC MEAN-ARITHMETIC MEAN INEQUALITY
There are many other refinements of (GA) of which we will mention
a few. THEOREM
24
[SIEGEL; HuNTER
J]
If n
> 2 and a a non-constant n-tuple then:
(a) (53)
where b is then-tuple defined by bk == 1 + (n- k)j(t + k- 1), 1 < k < n, and tis the unique positive root of n-2
~II(t+k)n+k-1 n! t- k k=1
II 1 3 and a constant.
0
The case n == 1 is trivial and for n == 2 see 2.2.1 Remark (i); so assume that
n
> 2. A simple application of 1.2(7) shows that the left-hand inequality in (57)
implies the right-hand inequality, and so it is sufficient to prove that
(58)
Means and Their Inequalities
151
The proof is by induction, and let x ==an, and then put the left-hand side of (58) equal to g(x ), and
f ==log og
Simple calculations show that f has a unique minimum at the point x == x' where (n- 1)2l (a)- Sj (a) x' == ;: ~ n - ; since n > 2 we have by (HA) that x' > 0. So for 1 all x =f. x',g(x) > g(x'). Simple calculations, and a further application of (HA) 1 show that g(x') > 2t~=~~~~t) (g) and so the result follows from the induction (5n-1 a
D
hypothesis. (vi)
REMARK
The result is in [Sierpinski]; this proof is from [Mitrinovic & Vasic
1976], and clearly gives the following Rado type extension of (58):
with equality if and only if a is constant. (vii)
REMARK
A simple proof giving the cases of equality can be found in [Alzer
1989a]; the same author gives a generalization of (57), [Alzer 1989b, 1991b].
An extension to weighted means, when the weights are decreasing, can be found in [Pecaric & Wang; Wang C L 1979e]. The following result generalizes these and inequality (57); [Alzer, Ando & Nakamura]. THEOREM
27 If a is a non-constant n-tuple, n
> 2, and w
is a decreasing n-tuple,
strictly if n == 2, define
If a is increasing then
f is strictly increasing on ] - oo, OJ, while if a is decreasing
f is strictly increasing on [0, oo[. In particular we get the following generalization of (58). COROLLARY
28 If a,
w are decreasing n-tuples, n
> 3,
then
with equality if and only if a is constant.
D
Take x == 0, 1 in Theorem 27.
REMARK
(viii)
D
It is clear that it is sufficient to require that the n-tuples a, w be
similarly ordered; see I 3.3 Definition 15.
152
Chapter II
REMARK
(ix)
The case, x
==
r, r
> 0, of Theorem 27 is given in III 6.4 Theorem
9. Daykin & Eliezer have given a convex function whose value increases from one side of (GA) to the other, thus providing many refinements of this inequality; [Daykin
& Eliezer 1967]. As these results follow from I 4.2 Theorem 18, we only give a later result of Chong K M. THEOREM
29
[CHONG
K M
1977]
If a is a non-constant n-tuple and
then A is strictly increasing. REMARK
(x)
This result generalizes (GA) since A(O)
==
Q5n(a; w) and A(1)
2tn(a; w).
This result has been extended in [Pecaric 1983-1984] where it is shown that
+ (1 - X)g_; w_) 2ln (X 2tn (a; w) + (1 - X) a; w) Q)n (X 2tn (g_; w_)
is an increasing function: and Chong K M has given another simple function with a similar property:
IT n
(
x 2tn (a; w)
+ (1 -
x) ai
)Wi/Wn
.
i=l REMARK
(xi)
THEOREM 30
For further results of this type see III 2.5.4. [WANG
C L 1980n] If a
a 0, k = 1, 2, ... ; see [DI pp. 71-72], where there are further references for these functions. THEOREM 8
If a and w are n-tuples, n
> 2, Wn == 1,
and if k E N*, tben
(5) witb equality if and only if a is constant. Tbe exponent on tbe left-hand side, k, cannot be replaced by any larger number.
D
The proof of (5) depends on showing that the function
f(x) == log (xk 11/J(k) (x) I), x > 0, is strictly convex, when the result follows from (J). For the rest if a is not constant and if (5) holds with exponent k replaced by a then,
< L~=l wdog l'lf{kl(ai)l-log 11/J(k) (2tn(Q;1!2.)) I a -
log 2tn (a; w) -log Q)n (a; w)
'
160
Chapter II
which, on letting a 1
-+
oo and using an asymptotic property of the multigamma
function, log l'l/J(k) (x) I rv -k log x, REMARK
(ii)
X-+
oo, implies that a < k.
0
For another converse inequality due to Alzer see below, 5.7 Theorem
18.
5 Some Miscellaneous Results In this section various properties of the elementary means of this chapter are discussed. The various results are not related to one another nor, in general, to the inequalities given earlier. 5.1 AN INDUCTIVE DEFINITION OF THE ARITHMETIC MEAN
We prove that the
equal weighted arithmetic mean of n-tuples can be defined from equal weighted arithmetic means of (n-1)-tuples. This fact has been used extensively by Aumann in his study of the axiomatics of means; see [Aumann 1933a,b]. THEOREM
1 Let a be an n-tuple and define the n-tuples aCr), r E N*, recursively
as follows:
aC 1 ) == the arithmetic means of then possible (n- 1)-tuples from a; aCr) == the arithmetic means of then possible (n- 1)-tuples from aCr- 1), r Then limr-+oo a~r) == 2ln(a), 1
0
> 2.
< i < n.
Assume for simplicity that for all r,
alr)
< a~). Then the result is immediate
from the following identities. n
n
n
L a~r) == L a~r-1) == L i=l
i=l
ai, r
> 2;
i=l
(r) _ (r) _ an al -
(r-1) an -
n-
(r-1)
a1 1
0 REMARK
This result has been extended; see [Kritikos 1949].
(i)
5.2 AN INVARIANCE PROPERTY
Given a sequence a it is natural to ask which
of its properties are also properties of the sequences of means, 2t, ®, defined in 3.4. We will only consider the sequence 2t in the equal weight case, and in several instances the answers are immediate: (a) if m
I:~ 1 aib~j), 1 < j < n.
Adding over j of these leads immediately to the required result. A similar proof can be given for (rv 1), when a and bare oppositely ordered.
D
Inequality (1), in the equal weight case, is due to Cebisev, and is called Cebisev's inequality; see [AI pp.36-37; DI pp.50-5; Hermite; HLP pp.43-44; PPT pp.197-
198],
[Herman, Kucera €1 Simsa pp.148-150, 159],
[Cebisev 1883; Daykin;
Djokovic 1964; Jensen 1888]. A history of this inequality can be found in the
excellent expository paper [Mitrinovic & Vasic 1974]. Generalizations are given in many places; see for instance [Pearce, Pecaric & Sunde; Pecaric 1985b; Pecaric & Dragomir 1990; Popoviciu 1959a; Toader 1996]. REMARK
(i)
The requirement that then-tuples be similarly ordered is sufficient
for ( 1) but it is not necessary; other sufficient conditions can be found in [Labutin 194 7]. The problem of giving necessary and sufficient conditions for the validity
of Cebisev's inequality has been solved; see [Sasser & Slater]. Nanjundiah has given a proof of (1) that yields a Rado type extension; [Bullen 1996b]. It is based on the use of Nanjundiah's inverse arithmetic means, see 3.4. 18
See I 3.3 Definition 15.
Means and Their Inequalities 5
LEMMA
163
With the notation of 3.4 Lemma 10 and assuming that n > 1 and
(an-1, an), (bn-1, bn) are similarly ordered
with equality only if an == an-1, or bn == bn-1·
0
This is an immediate consequence of the elementary computation
0 THEOREM
6 If n
> 1 and if a and b are both decreasing then
Wn(Qtn(a b; w)- Qtn(a; w)Qtn(b; w)
> Wn-1 with equality only if an
0
(Qtn-1 (a
== Q{n-1 (a; w)
b; W) -
Q{n-1 (a;
W)Qtn-1 (b; W)),
(3)
or bn == Q{n-1 (b; w).
Since a and b are decreasing so are 2l( a; w) and 2l(b; w), 5.2 (c). So we can
apply Lemma 5 using these sequences to get
by 3.4 Lemma 10 (a); which is just (3). The case of equality follows from that of Lemma 5.
0
A weaker result in the equal weight case, due to Janie, is that n 2 (2tn(ab)-
Qtn(a)2tn(b)) increases with n; [AI p.206], [Djokovic 1964]. The best result in this direction is the following, due to Alzer; [Alzer1989e]. THEOREM
7
If a, b are increasing n-tuples and w another n-tuple and k an integer
with 2 < k < n, and a1 < ak, b1 < bk then
REMARK
(ii)
The equal weight case of this result improves the quoted result of
Janie and the result of Nanjunduiah as it says that Q(n ( ab))
decreases with n.
(n 2 /(n- 1) (Qtn(a)2tn(b)-
164
Chapter II (i)
EXAMPLE
Suppose that
a2
== · · · ==an == b2 == · · · ==an == 1 then
2tn (Q Q; 312.) - 2tn (Q; 312.) 2tn (Q_; 312.) 2tk (a b; w) -2tk (a; w )2tk (b; w) WnWf(w1a1b1 + Wn- w1)- Wf(wlal WkW~(w1a1b1 + Wk- w1)- W~(w1a1
+ Wn- w1)(w1b1 + Wn- w1) + Wk- w1)(w1b1 + Wk- w1).
If now we let first a1 ~ 1, and then let b1 ~ 1 the right-hand side has limit Wf(Wn- w1) . ( ) • Th1s shows that the constant in the inequality in Theorem 7 2
wn wk -wl
cannot be improved. REMARK
McLaughlin & Metcalf have studied inequality (1) from the point
(iii)
of view of functions of an index set; see 3.2.2. They showed that under suitable conditions the difference between the right-and left-hand sides, considered as a function of index sets, is super-additive; [McLaughlin & Metcalf 1968b]. Another result due to Alzer is the following lower bound for the difference between the two sides in (1); [Alzer 1992c] THEOREM
If a and b are strictly increasing n-tuples and w another n-tuple,
8
n > 2, then
This inequality is strict unless for some positive a, {3, ai
bi == b1 D
+ (i
- 1) f3, 1
Let Cn
==
==
a1
+ (i -
1)a, and
< i < n.
min~i,j j,
.
.
2
(ai-aj)(bi-bj)==
2
L
(ak-ak-1)
L
(bm-bm-1)
k=j+l 2
L
(ak- ak-l)(bm- bm-1)
k,m=j+l 2
>
L
Cn == (i- j) 2 cn.
k,m=j+l Now if Sn denotes the left-hand side of the identity (2), the above calculation shows that
n
2 . Sn_ > Cn "'w·w·(i-J") L-t 2J 2 i,j=l
Simple manipulations show that this is just the inequality to be proved.
D
165
Means and Their Inequalities (iv)
REMARK
A related result can be found in [AI pp.340-341].
Writing (1) in the form
n
n
n (Lwiai)(~=wibi) < Wn(Lwiaibi) i=l
i=l
i=l
and applying this to infinite sums leads to various interesting elementary inequalities. For instance, [DI p.251]: tanxtany l og b - l oga > An(g). The similar proof of the second inequality also proves more, namely
55n (g) >
A (h) log b - log a b_ a > n- · D
In fact I 4.1 Remark (xiv) shows that 55n(a) and 55n(g) decrease with
(i)
REMARK
n while Qtn(g) and Qtn(h) increase, to the limits given above. COROLLARY
a+ b
-- > 2
e-1
13
1 ( bb) /(b-a) b- a > > v;;b aa logb-loga
(4)
1 log b -log a (ab) /(b-a) 2ab > a- 1 - b- 1 > e -ba > a+ b. REMARK
(ii)
N anjundiah has pointed out that this last inequality is a considerable
improvement on the standard estimates for e and the logarithmic function, in particular I 2.2(9). This is an implicit in a later proof of the same inequality by Kralik which consists of substituting 1 + x
== b/a in I 2.2(10). [Nanjundiah 1946;
KraJik]. 1
bb) 1/(b-a)
The quantities e- ( aa tric and logarithmic means 19
band b ~ are called respectively the iden1og - oga of a, b, written J (a, b) and .£(a, b); the definitions
are completed by putting J (a, a) == .£(a, a) == a. THEOREM
14 The means .£(a, b) and J(a, b) are strictly increasing as functions of
both a and b. D
This property of the logarithmic mean is a consequence of the strict con-
cavity of the logarithmic function and I 4.1 Remark(v). The same result and the strict convexity of the function x log x, see I 4.1 Example(i), gives the property for the identric mean. 19
The logarithmic mean has occurred earlier in 2.4.5 Footnote 15 and 4.1.
0
168
Chapter II
THEOREM
If a, b are positive numbers then
15
min {a, b} < ~ (a, b) < J (a, b) < max{a, b};
(5)
more precisely ~(a,
b)
2tj-1(Yn-j+1, · · ·, Yn-1),
< j < n,
is a convex function on [x1, Xn] and if A, B are defined by Ai
2ln-1(x~), Bj == 2ln-2(Y~) 1
< i < n, 1 < j < n- 1,
2ln ( f (X)) > 2ln -1 ( f (Y)) ; REMARK
2
(i)
then
2ln ( f (A)) < 2ln -1 (f (B)) ·
See [AI pp.233-234], [Bray; Popoviciu 1944; Toda]; see also V 2
Remark (x), and for another amusing result involving polynomials see [Klamkin
& Grosswald].
5.7
NANSON'S INEQUALITY
Means and Their Inequalities 17
THEOREM
If a is a convex (2n
+ 1)-tuple
and if b
171
== {a2, a4 , ... , a 2n}, c ==
{a I, a3, ... , a2n+ I} then
(9) with equality if and only if a is an arithmetic progression. 0
Since a is convex we have that ~ 2 an
> 0, I 3.1. So in particular for k ==
1,2, ... ,n k(n- k
+ 1)(a2k-1 -
2a2k
k(n- k)(a2k- 2a2k+l
+ a2k+t) > 0,
+ a2k+2) > 0.
Adding these inequalities gives (9). The case of equality follows since ~ 2 an
0
implies that a is an arithmetic sequence. REMARK
(i)
== 0
An equivalent result has been generalized by Adamovic & Pecaric;
see [AI pp.205-206; PPT pp.241-251], [Adamovic & Pecaric; Andrica, Ra§a & Toader; Milovanovic, Pecaric & Toader; Nanson; Steinig]. Another result involving convex n-tuples has been given by Alzer, [Alzer 1990d] THEOREM
18
If a is an n-tuple with the (n
+ 1 )-tuple 0, a I, ... , an
concave then
and the constant is best possible. 5.8
THE PSEUDO ARITHMETIC MEANS AND PSEUDO GEOMETRIC MEANS
If
a, w are n-tuples then an (a·w) _,_
==
W: 1 n ___!!:_ai- --~w·a· ~ ~ ~' WI
(10)
WI .
~=2
are called the pseudo arithmetic mean, and pseudo geometric mean of a with weight w, respectively. Other forms of ( 10) are worth noting:
(11)
aI
== Qln_,_ (a~· w) == Qj n_,_, (ab · w) ·
(12)
172
Chapter II
where -a~ and -ab are defined as:
REMARK
(i)
Simple examples show that in general the quantities defined in (10)
do not satisfy internality, see 1.1. This explains the term pseudo mean. Indeed if we assume that a1 == mina, then an(a;w) > mina == a1 implies, using (11), that a1
> 2tn-1 (ai_; wi_),
REMARK
(ii)
which is false.
This lack of internality is not surprising since these pseudo means
are cases of the arithmetic and geometric means with general weights that do not satisfy 3.7 (49), a condition that is necessary and sufficient for internality, see I 4.3 Remark( v); precisely, an(a; w)
The cases of equality follow from those of 3.1 Theorem 1.
0
Clearly Corollary 20 gives Rado-Popoviciu type inequalities for these pseudo means that are in a certain sense sharper than the original (R) REMARK
(vi)
and (P). As for (R) and (P) Corollary 20 implies the original inequality (13) and in fact can give better lower bounds, using the same idea as in 3.1 Corollary
REMARK
(vii)
3. In particular in the case of equal weights we get: g (a) - a (a) n -
n -
> max
2
r-+ex>
D
(a) Immediate
(b) We give several proofs of this result, and in all of them we assume that r E JR*, and without loss in generality that Wn == 1.
( i) [Paascbe] . lim log
r-+0
. log ( 2::~hm 9)1;[ ](a; w) == r-+0 r 2
-
1
Wiai)
(Wiai log ai) , -- r·1m L~=lLn r-+0
.
'1.=1
W,;ar: " 'l.
by l'H"'opita1' s R u 1e,
n
==
:L(Wi log ai) == 2tn (log( a); w), i=1
which gives the result by II 1.2 (8). The use of l'Hopital's Rule 1 is easily justified.
( ii) 1
log!JJtkl(a; w) = r log (
n
L wiai) i=1
using the Taylor expansion of the exponential function. The result is now immediate by I 2.2 (9), and II 1.2 (8).
(iii) By the note in the proof of I 2.1 Corollary 4 ai == 1 + bi, where bi == O(r) as r ~ 0, 1 < i < n; and by I 2.1(5) awir == (1 + bi)wi == 1 + wibi + O(r 2 ), as r ~ 0. Hence, ( rr~=l (1 + bi)Wi ) 1/r Q)n (Q; 1Q) L~ 1 Wi(1 + bi) vnkl(a; w)
1
== (
1 + 2::~= 1 w.;bi + O(r 1 + Li=l w'l.b'l.
== ( 1
+ 0 ( r 2 )) 1 I r
== 1
2
1 ))
+ 0 (r),
/r as r ~ 0,
L'Hopital's Rule is:if f,g are real differentiable functions defined on the interval ]a,b[, bounded or unbounded, with g' never zero and if, with c=a or c=b, either limx-+c f(x)=limx-+c g(x)=O or provided the right-hand side limits exists, finite =limx-+c gf;(c:z:)), limx-+c jg(x)l=oo, then limx-+c f~x~ X g X or infinite; [EMS pp.401-408].
Means and Their Inequalities
177
from which (b) is immediate.
( iv) Assume r > 0, then as in ( ii), log
mk1
1 (a; w) ==-log (
r
n
L. Wiar) ~=1
==log ~n(a; w). The opposite inequality follows from inequality (r;s) given below, 3.1.1. The case r < 0 is handled similarly. (c) Assume r E ~+ and without loss in generality that max a == an, then
which implies the first part of (c). A similar proof can be given for the other part of (c). A slightly different proof of the equal weight case is given below in 2.3. REMARK
(i)
D
It follows from the internality that if a is a strictly increasing se-
mtk1 (a; w), n EN*; see II 5.2 (c). More details of the behaviour of mk1
quence then so is REMARK
(ii)
as r -+ 0, ±oo can be found
in [Gustin 1950; Sbniad]. REMARK
(iii)
Another proof of the equal weight case of (b) is given below, 3.1.1
Remark (xi). The following simple identity is often useful. If r, s E ~*, t == s/r, b == ar, c == at then
(3) in particular of course, taking r == s in the first identity,
(4) REMARK
(iv)
These extend the identities II 1.2 (7), (8).
178
Chapter III
THEOREM 3 If a, u, v are n-tuples witb a increasing and Un == Vn, and if for 2
< k < n,
0
< Vn
- Vk-1
< Un
- Uk- 1 < Un
== Vn, tben for r
E
JR*,
Assume 0 < r < oo then
0
n
n
L uiar ==a1Un +
L(ui-1-
i=1
i=2
Un)~ai-1 n
n
>a1 Vn
+ LCVi-1- Vn)~ai-1
==
L viar. i=1
i=2
0 REMARK
(v)
In particular if 0
s > t > 0 then (8)
(c)
[RADON's INEQUALITY]
If p > 1 then
1 then
where the sup is over all b such that 2:::7 1 bf == 1; the sup is attained if and only if aP rv bp . (b) If p > 1 and a, b are n-tuples with An == Bn == 1 then I
I
i=l
with equality if and only if a REMARK
(vi)
rv
b.
The first part of this theorem is an extremely useful way of consid-
ering Theorem 1 and is the basis of a method of proof called quasi-linearization, see [BB p.23; MPF pp.669-619]. REMARK
(vii)
Many other inequalities can be considered this way; for instance
(GA): assuming without loss in generality that Wn
== 1,
n
QJn(a; w)
REMARK
(viii)
== inf
L WiCiai, where C == {c; QJn(c; w) == 1}.
cE 0 i=l
Proofs of (H) can be found in many places; see [AI p.SO; BB p.19;
HLP p.21], [Abou-Tair & Sulaiman 2000; Avram & Brown; Holder; Iwamoto; Liu
183
Means and Their Inequalities
& Wang; Lou; Matkowski 1991; Redbeffer 1981; Vasic & Keckic 1972; Wang C L 1977; You 1989a,b; Zorio]. 2.2 CAUCHY'S INEQUALITY
If p == 2, when of course p' == 2, (H) reduces to
(C) known variously as Cauchy's inequality, the Cauchy-Schwarz inequality 3 , or the [Bunyakovski~;
Cauchy-Schwarz-Bunyakovski1 inequality, see
we will refer to it
as (C). A discussion of the history of this inequality can be found in [Schreiber; Zhang, Bao & Fu]; an exhaustive survey of (C) and related inequalities can be found in [Dragomir 2003]. Obviously any proof of (H) provides a proof of (C). However, the simple nature of (C) allows for various direct proofs that show that 2.1 Theorem 1 holds for all real n-tuples a, b when p == 2. For instance, either of the following identities implies
(C):
n
n
L(aix + bi)
2
== x
2
n
L a7 + 2x L aibi + L b7 i=l
i=l
n
The second identity is known as the
i=l
i=l
Lagrange identity. For other proofs see
[1938a,b; Cannon; Dubeau 1990a,1991b; Eames; Sinnadurai 1963]. REMARK
rm
(i)
Of course (C) arises from (H) by taking p == p'. By taking
== · · · ==
== m in (2) we get a simple extension of (C); [Kim S]. n
THEOREM 5
D
r1
m
n
L a1j ... amj
m. Define
< i < m, 1 < j < n, < i < 21-L' 1 < j < n.
Then by (ii)
j=l i=l
j=l i=l
0; n 1 1rmina < R(r) < mina, r < 0.
This also shows that the equal weight case of 1 Theorem 2(c) can be deduced from the limit values of n( r). From the remark in the proof of I 2.1 Corollary 4 we have S(r) Hence, limr-tO Rn(r, a)
== n + O(r),
r ~ 0.
== oo, if n > 1. The same result follows from the simple
observations,
n 1 1rmina LEMMA
6
< R(r) < n 11rmaxa, r > 0; n 1 1rmaxa < R(r) < n 1 1rmina, r < 0. (a) R is strictly decreasing; that is if r < s then
R(s)
=
n
(
Lai
) 1j
8
0 then (11')
== 1. Then for all i, 1 < i < n, we have af < 1. So if s > 0, ai > af, showing that 1 < S(r), and so 1 < R(r); and if s < 0, ai < ai, showing that 1 > S(r), and so 1 < R(r). This completes the proof in this special case so now assume that S (s) == a, and put, for all i, bi == ai/a 1 1s == bi/R(s), when S(s;b) == 1. From the special case we then have that R(r; b) > 1, which is just (10). 0
(a) Assume that r > s and that S(s)
(b) First we consider S
The case of equality follows from that for 2.1 (7).
Means and Their Inequalities
187
Now we consider R . If R(r) > 1 then S(r) > 1 and so since logR(r) == r- 1 logS(r) the result follows from the convexity of logS and f(r) == r- 1, using I 4.1 Theorem 4(e). If R(r) == p < 1 choose p', 0 < p' < p and let a~ == ai/ p', 1 < i < n, when R( a', r) == pj p' > 1 and so R( a', r) is log-convex. However log R( a', r) == log R( r) - log p', so log R( r) is convex. D REMARK
(i)
REMARK
(ii)
The proof of (b) can be found in [Beckenbach 1946]. If w is an m-tuple with Wm == 1 a simple induction extends (11)
and (11') to S(~(r;w),a) R(~m(r;
< ®m(S(r1,a), ... ,S(rm,a);w),
(12)
w), a) < ®m(R(r1, a), ... , R(rm, a); w).
(12')
Ursell has made some interesting observations about (11) which we will now discuss; see [Reznick; Ursell]. Consider (11) in the special case obtained by putting r == 0 and As
== p,
s (p) < n 1-PIs SP Is (8 ).
(13)
or equivalently R(p) < nC!-~)R(s). Inequalities (10)-(13) are best possible in the sense that given n, r, s, A, p the constants cannot be improved.
However (10) and (13) are best possible in a
stronger sense. Given n, r, s,p, S(s) the values of S(r), S(p) are given by (10) and (13). In the case of (13) equality implies that (R(r)/R(s))rs/( 1-r) is equal ton, in particular it is an integer. In other words, given n, r, s, A and appropriate S (r) and S(s), we can still have strict inequality in (13) unless R(r)/R(s) has a special value. The determination of the exact range of this left-hand side given the sums on the right-hand side is completely determined by Ursell; see also [Pales 1990b] and IV 7.2.2. REMARK
(iii)
The sums, S and R, can be considered with weights; that is we
define S(r, a; w) ==
I:~ 1 Wiai and R(r, a; w)
==
(I:~ 1 Wiai) lr. It is immediate 1
that Lemma 6(b) remains valid for these more general functions. Lemma 6(a) is considered in [HLP p.29] and [Vasic & Pecaric 1980a]; it is valid if w
> e, while if
Wn < 1 the opposite inequality holds; see 3.1.1 Remark (ii). See also below IV 2 Theorem 13. (iv) Ifr > 1 then jR(r,a)-R(r,b)j <min{ I:~= 1 1ak-bikl}, where the minimum is taken over all permutations { i 1 , • . . , in} of { 1, ... , n}. This answers
REMARK
188
Chapter III
a question in [Mitrinovic & Adamovic]; some generalizations can be found in
[Milovanovic & Milovanovic 1978; Pecaric & Beesack 1986]. Because of (HA) and Lemma 6(a) the following strengthens (12'). LEMMA
7
IfWn == 1 then
R(55m(r; w), a) < 1 and r
n
1/
== A.p > p; then r' == A.q > q so
n
1/q
1. 1
af )
1/p
then f is strictly convex and homogeneous,
see I 4.6 Example (vii). So (M) is just I 4.6 (19). REMARK
(i)
to the case p
D
Of course basic properties of addition extends (M) in a trivial way
== 1.
Means and Their Inequalities
191
An important use of (M) is the proof of the triangle inequality, (T). If p > 1 then,
This inequality holds of course when p So if Pp (a, b)
==
( L~ 1
Iai -
bi IP )
== 1 being a property of the absolute value.
1/p
,p >
1, then we have the triangle inequality,
Pp(a, b) < Pp(a, c)+ pp(c, b),
(T)
the essential property for showing that pp is a metric on the space of n-tuples, or equivalently that
REMARK
(ii)
llaiiP ==
(
L~
1
af )
1/p
, p > 1, is a norm,
(M) was first used in the form (TN) by F. Riesz, and for this reason
Minkowski's inequality is sometimes called the Minkowski-Riesz inequality. For a similar reason the Holder inequality is sometimes called the Holder-Riesz inequality when written in the form
see [MPF pp.413--513], [Maligranda 2001]. The following extension of (M) is analogous to the extension of (H) given in 2.1 Corollary 2. COROLLARY
10 If ai
== (ai 1 , . . . , ain), 1 < i < m, and p > 1 then (16)
with equality if and only if then-tuples ai, 1 < i < m, are pairwise dependent. 0
Inequality (16), and so (M), can be proved by induction, as was the case for
(H), see 2.1 Corollary 2 proof ( ii). We start with the simplest case, the case m
==
n
== 2 of (16):
192
Chapter III
This can be obtained from 2.1(3), the case n
== 2 of (H), in the same way that
(M) was obtained from (H). Then, as with the inductive proof of 2.1 Corollary 2, we can either fix m
== 2 and give an inductive proof of (16) for all n, and then for
all m, or we can proceed the other way round; see [HLP p.38]. REMARK
(iii)
D
Much work has been done on this inequality; see for instance
[Szilard; Zorio]. 2.5 REFINEMENTS OF THE HOLDER, CAUCHY AND MINKOWSKI INEQUALITIES The inequalities (H), (C) and (M) have been subjected to considerable investigation, resulting in many refinements; some of these are taken up in this section. 2.5.1
A
The simplest proof of (H), or of 2.1 Corollary
RADO TYPE REFINEMENT
2, depends on (GA), and it is natural to ask if it is possible to refine (H) by using (R); [Bullen 1974]. THEOREM 11
With the hypotheses and notations of 2.1 Corollary 2 put n
~
(a)-
~
........ m -
then ifm
>2
L:j=l
1
1
Pm
Pm-1
with equality if and only if for j
m
Tii=l aij
) P1n
.' P7n/ri ri)
m (~n Ti i=l Dj=l aij
-Bm(a)
n
m
n
L II% >max{ 0, (1- ~m) II (L a~j) 1/ri }· j=l i=l
i=l j=l
D
The proof is based on the method used to obtain 2.1 Corollary 2 and Theorem 1. Use inequalities II 3.5(34), (35) applied to a sequence b, make the identification 2.5.1(18), and proceed as in the proof of 2.1 Corollary 2. REMARK
(i)
D
The cases of equality are discussed by both Kober and Diananda;
Diananda gives a similar refinement of (M); [Kober; Diananda 1963a,b]. 2.5.4 A
CONTINUUM OF EXTENSIONS
As has been remarked (C) is a particular
case of (H). In fact (H) can be used to prove much more. Let aij, ri, Pm be as in 2.1 Corollary 2; then we have the following result. THEOREM
14 If r
E
JR* and w is an n-tuple define the function f as follows m
f(x)
n
m
==II (L Wja:Jx/r II a~j x)l/ri. i=l j=l
k=l
195
Means and Their Inequalities
f unless f
Then
0
is log-convex, increasing on [0, oo[, and decreasing on ] - oo, OJ, strictly is constant.
If the m-tuples are not proportional then by 2.3 Lemma 6(b) and 2.3 Re-
mark(iii)
j=l
k=l
k=l
is log-convex, 1 < i < m. Hence, I 4.1 Theorem 4(d), n
log f (x) =
L :. log 9i (x) , .
~
~=1
1s convex. The rest of the proof is similar to that of I 4.2 Theorem 18; see [PPT pp.90 118], 7
0
[Mitrinovic & Pecaric 1987]. REMARK
(i)
Beside the paper just mentioned this result is discussed in [Daykin &
Eliezer 1968; Eliezer & Daykin; Eliezer & Mond; Flor; Godunova & Cebaevskaya; Mon, Sbeu & Wang 1992b]; see also IV 5 Example (xiv). REMARK
(ii)
== Pm
The most interesting case of Theorem 13 occurs on putting r
when f(O) is the left-hand side of a weighted form of 2.1(2), while f(Pm) is the
< f(x) < f(Pm), 0 < x < Pm, we get a continuum of generalizations of 2.1 (2), and so of (H): if 0 < x < Pm then right-hand side. Since f(O)
n
(
~ Wj
(
f! m
)
aij
Prn) 1/ Prn
o, i=1
i=l
i=1
i=l
which is an immediate consequence of (C). REMARK
(iv)
D
If x == 0 then (21) reduces to (C).
(v)
This result is due to S.S. Wagner, but the above simple proof is by Flor; [MPF p.BS], [Andreescu, Andrica & Drimbe; Flor; Wagner S]. REMARK
2.5.5 BECKENBACH's INEQUALITIES
The first part of 2.1 Theorem 1 can be inter-
preted as follows. and b then (H) holds for all a'I!b'I! == a'f!bP == aPb'f! 2 < i < n. ~~ ~1 1~,-Given
{t1
I
I
a2, ... ,
an, with equality if and only if
I
This has been generalized as follows in [Beckenbach 1966]. THEOREM
a as
16
Let a and b ben-tuples, p > 1, and 1 < m < n; define then-tuple
follows: if 1
< i < m,
Means and Their Inequalities
197
then given a1, ... , am and b,
(22) for all am+ 1 , ... , an .
=f.
If p < 1,p ai == iii, m
0
0, (rv22) holds.
There is equality in both cases if and only if
+ 1 < i < n.
Assume that p > 1 and consider the left-hand side of (22) as function
the variables am+1, ... , an. Then
fj
== PQj,
m
f of
+ 1 < j < n, where
n
n
Qj == (Laibi)a~- - (Laf)bJ. 1
i=l
Since P > 0 we have that the partial derivatives
are zero; that is if and only if
~~.
a~. =f:i:=l~r.' i=l a~ z
fj
are zero if and only if the Qj
= fJ= 1 ar., m + 1 < j < n. Equivalently, if i=l a~ ~
aJ J
andonlyif
i=l
m+l<j 0, f3 > 0, ry > 0, and a, b are non-negative
THEOREM
n-tuples; for m, 0 m
+ 1, ... , n.
+
with equality if and only if ai == ci, m
(a+ 'Y I:~
1 m+1 /P --------~------ry I:~ m+l bici
cf)
f3 +
+ 1 < i < n.
0 n
n
i=m+1
i=m+1
< a+ry ~ ~ af
(
i=m+1
)1/p(a-P'/Pf3P +rr 1
~
~ bf i=m+1
')1/p' ,
by (H),
198
Chapter III
The case of equality follows from that of (H).
D
(iii)
A converse for this inequality has been given by Zhuang; [Zhuang
REMARK ( iv)
All of these inequalities have been given considerable attention;
REMARK
1993].
[Iwamoto & Wang; Wang C L 1977,1979b,d,1981a,1982,1983,1988b]. 2.5.6 OsTROWSKI's INEQUALITY
This inequality is an extension of (C); [Ostrowski
p.289]. THEOREM
18 Let a and b be n tuples such that a rf b and define the n-tuple c
by a. c == 0 and b. c == 1, then
with equality if and only if
REMARK
(i)
A proof can be found in [AI pp. 66-70], where several extensions
due to Ky Fan & Todd are also proved; [Fan & Todd]; see also [Madevski; Mitrovic 1973] REMARK
(ii)
Beesack has pointed out that this result can be regarded as a special
case of a Bessel inequality for non-orthonormal vectors; [Beesack 1975]; see also [Sikic & Sikic]. It can also be regarded as an extension of 3.1.2 (7) in the case
r ==
q
== 2, s == 1.
2.5. 7 ACZEL-LORENTZ INEQUALITIES
As was pointed out in 2.4 (M) is essential for
certain properties of norms on JRn. If a different norm is defined then we can ask if similar properties hold. A so-called Lorentz norm 4 is defined as n
\\a\\~== (af- Laf) 11P, p > 1; i=2
this is defined on the set Lp of positive n-tuples for which
a1
>
(2:~
2
af) 1 1P; [BB
pp.38-39]. Various inequalities using these expressions are then called Lorentz
inequalities, [Wang C L 1984a]. However the first such inequality was proved by 4
This is H.A.Lorentz; see [EM6 p.46-47].
199
Means and Their Inequalities
Aczel, [Aczel 1956b], so such inequalities have also been called Aczel inequalities; [DI pp.16-17, PPT pp.124-126].
It turns out that the natural analogue in this situation is not (M) but (rv M). As a result of this the Lorentz triangle inequalities are analogous to (rvT), (rvT N), and so the Lorentz norm is not a norm in the usual sense. The various Aczel-Lorentz inequalities are examples of reverse inequalities similar to those in I 4.4. THEOREM
where p
19 (a)
If a, b are n-tuples in Lp, Lp' respectively,
[HoLDER-LORENTz]
> 1 then (23) i=1
i=1
i=1 I
IfO
< p < 1 then (rv23) holds; in either case there is equality if and only if aP rv bP .
(b) [MINKowsKI-LORENTz] If a, bare n-tuples in Lp, p n
> 1,
then
n
n
((a1 + b1)P- L(ai + bi)P)l/p > (af- Laf) 11P + (bf- Lbf) 11P. i=1
IfO
(24)
i=1
< p < 1 then (rv24) holds, and in either case there is equality if and only if a
and b are dependent. (c) [BECKENBACH-LORENTz] Suppose p non-negative n-tuples; for m, 0 (abi/ f3)P' IP, i a - 1 L~
cf > 0. (a -
n, define the (n- m)-tuple c by ci m+ 1
are
==
af > 0, and that
Then I
I:~=m+1 af) ljp > (a - I L~=m+1 cf) l/p
f3 - I L~
m+1
with equality if and only if ai REMARK
< m
1, m < n, then
Bn- Bm-k-1 + ( (k + l)liP )
p')
1I
P'
'
where k is given by (25). REMARK
An extension of this result has been given in [Iwamoto, Tomkins &
(i)
Wang 1986b], and a simpler proof of this extension, using Steffensen's inequality,
VI 1.3.6 Theorem 18, can be found in [Pearce & Pecaric 1995a]. The following result is in [Mikolc:ls]; it reduces to (M) if pr
== 1; see [AI
p.369],
[Alzer 1991i]. THEOREM
21 If aj, 1 < j
< n, are m-tuples and 0 < r < 1, 0 < pr < 1,
then
(26) Ifr
> 1 and pr > 1
then (rv26) holds.
The next result is a problem set by Klamkin, [Klamkin & Hashway].
It is a
particular case of Callebaut's inequality, 2.5.4 Remark(iii). THEOREM
22 If a and b are n-tuples and if
f (k)
==
~ ai (~
2-k
i=l
then
f
n
.!£_) (~ 1£_ 2-1£_) b; ~a; bi n ' 0 < k < n; i=l
is decreasing.
REMARK
The extreme inequality given by Theorem 22, f(O) > f(n), is just
(ii)
(C) The following is due to Abramovich; [Abramovich]. 23 If p > 1 and a, b, c are n-tuples with An == Cn == 1 and for some Ci Cj 1 m < n, ai >c.;, aj < Cj, bi > bj, < z. < m, m + 1 < J. < n, t h en
THEOREM
m, 1
0, r =I- 0, from which (r;s) is immediate.
This proof can be given without appealing to (C), [Burrows & Talbot; Wahlund]. (iii) Make the same assumptions as in proof ( i). Define then-tuple b by b =
2tn(:; ill).
Then it is sufficient to prove that
(4) From the definition of b, L~ f3i > -1 and I::~ 1 wif3i Hence, using (B),
Lr
1
wibi == 1, so if we put bi == 1 + f3i, 1
== 0. n
1
< i < n, then
Wibf ==
n
L wi(1 + f3i) > L wi(1 + sf3i) == 1. 8
i=1
i=1
This immediately gives (4). See [Herman, Kucera €3 Simsa pp.161-168]. ( iv) [Orts] Assume that a is not constant.
(i) First assume that r, s E N*.
m~l (a; w) < m~J (a; w) follows by (C). Now assume that m~l(a;w) < Wl~+l](a;w), 1 < k < m -1. Then
< ( m~m+ 1 ] (a; w) ) m+1 ( 9)1~] (a; w) )m-1 'by the induction hypothesis. This gives 9)1~] (a; w)
< m~+ 1 ] (a; w), and completes the proof for this case.
Means and Their Inequalities
205
(ii) Now suppose that r, s E Q, and that r == yjx, s == zjx where x, y, z E N* and y < z; then
!JJt[l(a; w)
=(i:: wi(aYx?)
l/r
i=l
(iii) The general case of positive real r, s follows by taking limits, and as can easily be verified strict inequality is obtained.
(v) Assume that Wn == 1 and consider the extreme values of the function 2::~ 1 wiaf as a function of a subject to the condition 2::~ 1 Wiai == A we easily see that if s > 1 then 2tn(a; 717) < 9Jtk1 (a; w ), while if s < 1 the opposite inequality holds. (vi) We now give an inductive proof of the equal weight case of (1;s) that has been attributed to Steinitz; see [Paasche]. In the case n == 2 assume that 0 < a < b when we have to show that 2((a + b)/2)s 0 and the property of I 4.1 Lemma 2. Alternatively, putting {3 == (b- a)/2, (5) can be written as
(b- f3)s - (b - 2{3)s < b - (b - {3) 8
== y
(y - {3) 8 , y > {3, is strictly increasing, the last inequality being just g(b - {3) < g(b).
which follows by noting that the function g(y) Now let a be a non-constant n-tuple, n
8
-
> 3, and without loss in generality assume
== an, min a == an- when by the strict internality of the arithmetic mean an- < 2tn(a) 0, and so mtk] (a; w) is a convex function of 1/r, r > 0. This last result was pointed out by Beesack, who used it to give a simple proof of the following result due Hsu; see [Julia], [Beesack 1972; Hsu; Liapunov; Rabmail 1976]. THEOREM
2 If 0 < r < s < t and a is not constant tben 1
< 9Jt~l (g; w.) - 9Jtkl (g; w.) < s(t - r). 911~] (a; w) - mtkl (a; w) r( t - s) '
andifO:i a: Y) ~\ 1
2
E N*, we get an increasing sequence with
limit an_ 1 an; equivalently 1/r
lim
r-+oo
Means and Their Inequalities
211
In general 1/r
lim
(6)
(n- P + 1) L !(I:~-: aii)r
r~cx::>
p-l
The values of the numerator in (6) are easy to obtain if we take r == 2, 4, 8, .... This is done by forming equations whose roots are the sequences of those of the original equation, and iterating this process. If the k-th equation by this procedure is xn
+ c~k) xn- 1 + · · · + c;;~ 1 x + c~k)
== 0, and then the ratio in (6) is just
In this way, by forming all such expressions, every root of the original equation can be approximated simultaneously. A different kind of generalization of Theorem 1 is the following result which contains the equal weight (r;s) as a special case; see [Marshall, Olkin & Proschan]. THEOREM
then
3 If a and b are n-tuples such that b is decreasing, but b/ a is increasing
(I:~ ai /I:~ bi) 1
1
1
/r increases with r
Various rewritings of (r;s) can be found in [Kim Y 2000b]. 3.1.2 HoLDER's INEQUALITY AGAIN
In this section we will consider the inequality (7)
Similar inequalities for power sums have been considered earlier; see 2.2 Remark (i), 2.3 Corollary 8. It is convenient to deal with various simple cases of this inequality and then state a theorem that will cover the remaining more interesting situations. As there is no loss in generality we will assume that q
< r,
q, r E JR. Further since
(7) is an identity if both a and b are constant we assume that this is not the case. (a) If either a or b is constant then (7) reduces to a case of (r;s). For instance, if
a is constant, when by the above assumption b is not constant, then (7) holds as (s;q) if
8
< q' strictly if 8 < q' while (rv7) holds if 8 > q' strictly if 8 > q.
(b) If q == r == oo then (7) holds by I 2. 2 ( 16) and (r;s), strictly unless for some i, 1
< i < n, ai
(r;s), strictly unless
== oo strictly if r < 8. (c) If q < r
8
8
== oo and
==max a, bi ==max b. If q == r == -oo then (rv7) holds by
== oo and for some i, 1
then (7) holds strictly if
8
< i < n, ai
0.
== 0 then (rv7) holds for all s > 0, strictly if s > 0 or if s == 0 and q < 0. (f) If q == r == s == 0 then (7) is an identity. (e) If r
4 If a, band w are n-tuples, a b not constant, then (7) holds if q, r, s E ~* and satisfy either (a) 0 < q < r and 1lq + llr < 1ls, or (b) q < 0 < r and 1lq + 1lr < 1ls < 0. If q < 0 < r and 1lq + 1lr > 1ls > 0, or q < r < 0 and 1lq + 1lr > 11 s then (rv7) holds. The inequality is strict in both cases unless 1Iq + 1Ir == 1Is and aq rv br . THEOREM
D
Let
In (a) t >
~ + ~ == ~
q r 0 and s
t
then
!q +!r == 1 so tlq and tlr are conjugate indices.
< t so
9Jtkl(ab;w) < 9J1Kl(ab;w) ==(2tn((ab)t;w)) 11 t, by(r;s) and 1 (4) 0, 1 < i < m, 1ls > 2:::~ 1 1/ri and ifa(i), 1 < i < m,w, are n-tuples with IT~ 1 aCi) not constant then COROLLARY
m
m
i=l
i=1
m~l(II a 0, 1 < i
inf_~Ec I:r wiciai + inffEc :z=r wicibi, by I 2.2 Q5n(a + b; w) == inf
cE 0 i=1
1
1
1
(e) (rv 15) so we get (10) the case r
== 0 of (a).
More simply: with c ==a orb and using (GA),
Q5n(c;w) (g + Q; 1!2.)
Q) n
= ®n (C/( a + b); W ) < 2tn (c/( a + b) ; w ) ;
now add the two inequalities obtained by taking the two values of c. The case of equality follows from that of (GA).
(ii) Since by I 4.6 Example (viii) the function x(a) == Q5n(a; w) is strictly concave (10) follows from II 4.6 (rvl9); see [Soloviov]. (b) Assume r, s E ~' rs i= 0, r < s, and without loss in generality that Um == Vn == 1 when (9) is
or
but this is just 2.4(16). The other cases are proved by taking limits. REMARK
D
(i) The cases of equality in (9) are discussed in [HLP p.31]. This in-
equality is sometimes called Jessen's inequality, [MPF p.108]; it has been much generalized, see [Toader 1987b; Toader & Dragomir]. (ii)
These results have been proved by many authors; see [BB p.26], [Beckenbach 1942; Besso; Bienayme; Giaccardi 1955; Jessen 1931a; Liapunov; Norris 1935; Schlomilch 1858b; Simon].
REMARK
(iii)
Inequality (10) is sometimes called Holder's inequality. A simple proof of (10) can be given using the inverse geometric means of Nanjun-
REMARK
diah, II 3.4. The proof in fact gives a Rado-Popoviciu type extension of (10).
215
Means and Their Inequalities THEOREM
0
8 If n > 1 then
By II 3.4 (22) and II 3.4 Lemma 10 (a) Q;n 1 (Q5(a; w)+Q;(b; w); w)
< Q;n 1 (Q5(a; w); w) + Q;n 1 (Q5(b; w); w) == an
+ bn == Q; n 1 ( Q; (a + b; w) ; w),
which gives the above inequality. The case of equality follows from that of II 3.4 Lemma 10 (c). REMARK
(iv)
0
A simple proof of the right-hand inequality of II 2.5.3(36) can be
given using (10).
+ Q; n (a; W) < Q; n ( e + a; w), by 0,
1
> 1 then
< j, k < n.
A direct proof of this inequality using (H) can be found in [Bullen 1972]. 3.2.3 EXTENSIONS OF RADO-POPOVICIU TYPE
An extension of (R) and (P) to power means is given by the following result.
Assume that a and w are n-tuples, n (a) If r < s and A == r or s but A -/= 0, then
THEOREM
13
> 2, and r, s
E ~'
r f=. s.
(12)
with equality when A == s if and only if an == Wt~~ 1 (a; w), and when A == r if and only if an== 9Jt~~ 1 (a; w). (b) If r
< 0 < s then
w))
9Jt[s] n (a· _,_ (
9J1kl (a; w)
Wn
>
(13)
Chapter III
218
with equality, when s == 0, if and only if an== 9.Rt~ 1 (a;w), when r == 0, if and only if an== 9Jt~~ 1 (a; w), and when r < 0 < s, if and only if both conditions hold. (a) The case r
0
=I- 0, s == oo, when A== r, or s =I- 0, r == -oo, when A== s are
almost immediate. As in 3.1 Theorem 1 we can if s
=I- 0 use 1( 4) to assume s == 1, or if r =I- 0 that
==
1. Further if either s == 0, orr== 0, then (12) reduces to (R). So it suffices to consider the cases (i) r == 1 < s < oo, and (ii) -oo < r < 1 == s, r =f- 0. r
Case (i) In this case if
A== s then (12) is equivalent to
n-1
n
i=1
i=1
w~=:(L Wiair + w~-"(wnanr > w~-s(L Wiair, and when A == 1 to +w1-1/s(w as)1/s w1-1/s(~w-a~)1/s > w1-1/s(~w-a~)1/s · n n n L.._.,; n-1 L....,; n ~
~
~
~
i=I
i=1
An application of (B), or (H), confirms these inequalities and gives the cases of equality. Case (ii) To consider this case put x
== an
assume that .X
== s == 1, and put
f(x) == Qtn(a; w)- wtk] (a; w). Then ( . ) Wn ) _ ( Wn-1 Wn mn-1 a,w + Wn X
f (X ) -
-
( Wn-1 ( [r] ( . )) r Wn r) l/r ' + Wnx Wn 9Jtn-I a,w
and
f has a unique minimum at X== Xo == mk~1 (a; w). As a result f(x) > f(xo) if x =I- x 0 , and simple calculations show that this is the
So
desired result. (b) The cases of non-finite r and, or, s are straightforward. Further, if r, s E ~ and either r or s is zero the result reduces to (P) by using 1 (4). So we may assume
Means and Their Inequalities
219
that r, s E JR*. Then
1og ( 9]l~l(g;1Q))Wn [r] . Wln (a, w)
W
(!l (Wn_ n s og W n
1 (
W
1
. ~) + W ans)
~
~ w~a~
n-1 i=1
Wn
n
The case of equality follows since the logarithmic function is strictly concave. REMARK
(i)
D
Another proof of one case of part (a) of this theorem is given below,
see 3.2.6 Theorem 24. REMARK
(ii)
The technique used in II 4.1 Theorem 1 proof (i) can be applied to
the proof of (a) case (ii) to obtain inequalities converse to (r;s). This will not be done here as the whole topic will be taken up in 4 below. REMARK
(iii)
Many similar results follow from more general theorems to be
proved later, see IV 3.2, and so we will not discuss them at this point; see [Bullen
1968; Mitrinovic & Vasic 1966a,b,c], A simple deduction from Theorem 13 is the following. COROLLARY
14 If a and w are n-tuples, n
> 2, and
-oo
< r < 1 < s < oo, r #- s
then
with equality when s == 1 if and only if an == mk~ 1 (a; w), when r == 1 if and only
if an == mk~ 1 (a; w), and when r
< 1 < s if and only if both
of these conditions
hold. D
Take A == r == 1 and A == s == 1 in (12) and add the resulting inequalities.
The cases of equality are immediate. REMARK
(iv)
(R) and (P).
D
Inequality (14) and its analogue (13) are the simplest extensions of
220
Chapter III
If we do not assume that r
< 1 < s then (14) need not hold as the following
example of Diananda shows. Assume that 1 < r < s < oo, 0 < 8 < 1, w1 == · · · == Wn == 1, n > 3, · · · == an-s ==an== 1, an-2 == (1+6) 118 , an-1 == (1-8) 1/s. Then (14) reduces (i)
ExAMPLE
a1
==
to 1 + (n -1)mk~l(a;w) > nmk1(a;w) or equivalently 2ln(b) > mt](b), where then-tuple b == (k, ... , k, 1), with k == (n -1) 1/T(n- 3 + (1 + 8)T/s + (1- 8)Tis). Since k an- 2 ,
#
1 this is false by (r;s). If r < s < 1 repeat the above replacing s by r in
and in
an- 1 ·
We now prove a general theorem that implies many extensions of the results in II 3.2.2 ; see [Mitrinovic & Vasic 1968d]. 3.2.4
INDEX SET EXTENSIONS
Let A, J-L > 0, A+ J-L > 1, a, b, u, v sequences, I1, I2, J1, J2 index sets with I1, J1 disjoint, I2, J2 disjoint, then
THEOREM
UA
11UJ1
15
VJ.L
12UJ2
(a·u))AT(wt[s]
(wt[T]
11UJ1 - ' -
(b·v))J.LS
>U.A VJ.L (m[T](a· u))AT(m[s](b· v))J.LS -
11
I2
I1 - ' -
(15)
12UJ2 - ' -
12 - ' -
+ UAJ1 vt-tJ2 (VR[T](a· u))AT(mt[s](b· v))J.LS. J1 - ' J2 - ' -
If A+ J-L > 1 (15) is strict; if A+ J-L == 1 (15) is strict unless U11 AVI-L (vn[T] (a· u)) AT (vn[s] (b· v)) J.Ls == UA VI-L (vn[TJ (a· u)) AT (vn[s] (b· v)) t-ts. J2 11 _,_ J2 - ' J1 12
J1 - ' -
I2 - ' -
(16)
If AJ-L < 0, A + 1-" == 1 then (rv 15) holds, with the same conditions for equality. 0 A
==
This follows immediately from 2.3 Corollary 8(b) with n == 2 on putting 1/p' f""'11. == 1/p' and aP1 == UI1 (vn[T] (a· u))T ' aP2 == UJ 1 (m[T] (a· u))T ·' and bp' == 11 - ' J1 -'1
v12 ( m}:] ( b;
v)) s' b~
1
== VJ2 ( vn~; (b; v)) s. The cases of equality are immediate.
0
(i)
Since Theorem 15 is such an immediate consequence of a very simple case of (H) any deduction from this theorem can also be obtained directly and REMARK
easily from (H) . REMARK
(ii)
Inequality (15) is easily extended to allow for m pairs of disjoint
index sets. REMARK
(iii)
Put I1
== J2 == I, J1 == J2
== J,
r == p, s == p', A == 1/p, J-L == 1/p',
ui == 1 ==vi, i EN* then (15) becomes 1 11 11 p ( /p' > p
( L af) iEIUJ
L bf')
(L af) (L bf')
iE1UJ
iE1
11 p'
iEI
+ (L af) iEJ
11 p
(L bf')
1
/p'.
iEJ
Since clearly
L aibi == L aibi + L aibi IUJ
I
J
the last inequality implies the first part of 2.5.2 Theorem 12, and the case of equality in that theorem is an easy consequence of that in 2.4 Corollary 10.
Means and Their Inequalities 16 If I, J are disjoint index sets, a, u, v are sequences, r, s E
COROLLARY
0
221 r
-
(wt[r] (a·
usl(s-r) I
I
u))
rsl(s-r)
+
-' -
0'\'l-[s] ( . ) ;.J.Jl-I a,v
vrl(s-r) I
(wt[r] (a·
usl(s-r) J
J
u))
rsl(s-r)
-' -
.
0'\'l-[s] ( . ) '..J-'l-J a,v
vrl(s-r) J
The inequality is strict unless
If r s > 0, r -/:- s (rv 17) holds and is strict under the same conditions.
D
This is an immediate consequence of Theorem 15, by taking a == b, I 1
I2 ==I, J1 == J2 == J, A== s/(s- r), M == -r/(s- r). REMARK
(iv)
Defining the following function on the index sets
then Corollary 16 says that v is super-additive if r s REMARK
(v)
< 0 and sub-additive if r s > 0.
Taking I== {1, 2, ... , n- 1}, J == {n} (17) becomes
( . ) ) rsl(s-r) U nsl(s-r) (0'\'l-[r] '..J-'l-n Q, 1k v: l(s-r)
D
wt~1 (a; v)
> -
0'\'l-[ r] ( . ) ) r s I ( s- r) ;.J.Jl-n-1 Q, Y:.. ( . ) v:~~s-r) ( 0'\'l-[s] '..J-'l-n-1 a, V
( r) + _u_n1_ _
usl(s-r) n-1
8
8_
rl(s-r)'
Vn
(18) with equality if and only if
(
REMARK
(vi)
Un-1 Vn)
a~-r = (9J1k~1 (g; .Y.))
Un Vn-1
8
(wt[u~ (a· u))u. n
1
-'-
Inequalities of this type for power means seem to have been first
discussed by McLaughlin & Metcalf, and independently by Bullen; see [Bullen 1968; McLaughlin & Metcalf 1967a,b,c].
222
Chapter III
(19) Inequality (19) is strict unless mtr] (a; u)
== mt~] (a; u).
The right-hand side of (17), divided by UiuJ can be written in the form
D
[r] ) rj(s-r) UI UI (9Jti (g;1!!_))s UiuJ ( VI VR~s] (a; v)
+
UJ [!JuJ
(
(r]
UJ (9JtJ (Q;lf))s VJ 9)1~] (a; v)
)
rj(s-r) '
which by (GA) is greater than or equal to
UI (9J1f] (g; 1!!.)) s) rUr /(s-r)UruJ ( U J (wt~) (g; 1!!.)) 8 ) rUJ j(s-r)UruJ ( VI
mYJ(a;v)
wt~](a;v)
VJ
On taking suitable powers of this inequality (19) follows. REMARK
Taking I == { 1, ... , n}, J
(vii)
then letting r
-t
D
== {n + 1, ... , n + m}, u == v, s == 1 and
0-, (19) reduces to II 3.2.2 Theorem 8(b), together with the case
of equality, making Corollary 17 a generalization of that result. Define the following function on the index sets, a(I)
== Wiwt}sJ (a; w), then the
following is an analogue of 2.5.2 Theorem 12; [Everitt 1963]. COROLLARY
18
If s
> 1 and
I, J are disjoint index sets then
a(I U J) > a(I) with equalityifand onlyif9J1~](a;w) with the same case of equality. If s
(20)
== 9J1~J(a;w). Ifs < 1 then (rv 20) holds
== 1 then (20) is an identity.
The case s == 1 is trivial so suppose first that s > 1.
D
In Theorem 15 take I1
u
+ a(J),
== I2 == I, J1 == J2 == J, r == 0, .A == 1 -
J-L, J-l == 1/ s, and
== v == w; the result is then immediate.
If s < 1, s =/= 0 there is a similar proof. If s
== 0 the right-hand side of (20) is 1
WiuJ(:
IUJ
QJI(a;w)+:J QJJ(a;w)) IUJ
<WiuJ(!)I(a;w)Wr/WruJ(!)J(a;w) ) ==WJuJ(!) ruJ(a; w). The case of equality follows from that of (GA).
WJ/WruJ
, by (GA),
D
The following generalizes II 3.2.2 Theorem 8 and the notation used is introduced there; see [Bullen 1968; McLaughlin & Metcalf 1967b, 1968b; Mitrinovic & Vasic,
1968d].
Means and Their Inequalities THEOREM
223
19 Ifa,u and v are (n+m)-tuples, A E R, 0 < Ar/(s-r)Um < Un+l,
sjr > 1 then Vn+m (wt[s] (a· v)) s - [fn+m (wt[r] (a· u)) s Vm n+m _,_ Um n+m _,_ ~
(21)
s
[Ts/r s > n (wt[s] (a· v)) - n (wt[r] (a· u)) (U - U Ar/(s-r)) (r--s)/r - V n -' U n -' n+m m m
m
+ (DR~(a;v)r- ~ (DR~(a;u)r; and ~ s U s 9Jt[s] ( ) s n+m (wt[s] (a· v)) -A n+m (wt[r] (a· u)) ( m Q; 12. ) V n+m _,_ U n+m -'-[r] m
wtm (a; u)
m
>
-vVnm
(wt[sl(a·v))sn -'-
(22)
u~lr (wt[rl(a·u))s(U U n _,_ n+m
U
m
Ar/(s-r))(r-s)/r(9R~(g;Q))s -[r] wtm (a; u)
m
< sjr < 1 then (rv 21) and (rv22) hold. Equality occurs if s == 0, when no restriction need be placed on A, or if s =f. 0 if un m[r]n (a·_, _u) == (Ar /(r-s) [1.n +m - um )l/r wt[r]m (a·_, _u). IfO
Letting r
-+
0 in
(21), (21), gives inequalities that hold for all s, that are strict
except under the conditions obtained from the above conditions on letting r
D
Proof of (21) case (i):r =f. O,s=f. 0.
to
-+
0.
In this case (21) is equivalent
n+m n+m 1 """" . 8 Un+m ("""" . r) s/r =~ v~ai~ u~ai Vm.~=1 Um . ~=1 n
> _1 """" . ~ __ ~ v~a~ Vm.~=1
n
_- _ 1 ("' . r:) sjr (Un+m _ U mA\ rj(s-r)) (r-s)/r ~ u~a~ Um ~=1 .
This, in turn, is equivalent to n
' . r)s/r(U _ U ,rj(s-r))(r-s)/r (" ~ u~ai n+m mA
(23)
i=1
+
n+m \Tj(s-r))(r-s)jr("' (u mA ~ i=n+1
n+m . r:)sjr > u(r-s)jr(" . r:)s/r. u~a~ _ n+m ~ u~a~ , i=1
and this last inequality follows from the n == 2 case of (H), as does the case of equality.
224
Chapter III
Proof of (21) case (ii):r#O,s==O.
Now (21) is equivalent to
+ 1 + A-1
Vn+m _ Un+m > Vn _ 1 (U _ U A-1) Vm Um - V m U m n+m m
'
which reduces to an identity for all A. Equivalently we could use (23). Proof of (21) case (iii) : r == 0, s =I= 0. Letting r --+ 0 (21) becomes
But (24) is equivalent to
or to Um Un+m TT
(Qj m (a·_,-u))s + UUn (Q) n (a·_, _u))s AUn+rniUn >_ A(Qj n+m (a·_, _u))s , n+m
which is a special case of (GA). The case of equality follows from that of (GA). Proof of ( 20) case ( i v) r == 0, s == 0. This case is covered by the previous case where the assumption s =/=- 0 is not used. Proof of (22)
In all four cases (22) holds under the same circumstances as (21); we see this as follows in two important cases. Case r =/=- 0, s =/=- 0. After some simplification (22) is seen to be equivalent to (21). Case r == 0, s =/=- 0. Letting r --+ 0 (22) becomes:
Vn+m Vm
(m[s]n+m (a·_,_v)) s >
Vn - Vm
A Un+m U
m
(Q) n+m (a·-'-u)) s
(m[s]n (a·_,_v)) s -A Un+rn/U rn Um Un (Qj
-[s]
9Jtm (g_; 1!.)
(
t1t
(
) s
•
vm a,v -[s]
(a· u)) s 9Jtm (g_; 1!.)
n _,_
(
which after some simplification is seen to be equivalent to (25).
QJm(a;v
) s
. D
Means and Their Inequalities REMARK
(viii)
REMARK
(ix)
225
It should be remarked that neither (21) nor (22) depend on v. If we take .-\
== 1, u == v then (22) shows that if s/r > 1 then
is super-additive, which generalizes II 3.2.2 Theorem 7. Included as special cases of these results are the inequalities in II 3.2.1, in particular those of Example (iii); and for further results see below 3.2.5 and IV 3.2. 3.2.5 THE LIMIT THEOREM OF EVERITT
The result of II 3.3 is easily extended to
power means as was pointed out by [Everitt 1967,1969]. His results are included in the following theorem, [Diananda 1973.] THEOREM
s
Wn-t((9J1~~ 1 (a;w)) - (9Jt~~Ja;w)) ), 8
8
(27)
with equality if and only if an == 9J1~~ 1 (a; w). D
This follows as in proof (v) of II 3.1 Theorem 1, but using (26) and noting
that Wn(wtk1)s, n EN, is increasing;
which is what we have to prove. The case of equality follows from that for Lemma 23(b).
0
We can also give an analogue for (M), and use it to prove a Rado type extension of (M).
Means and Their Inequalities THEOREM
25 Ifr > 1 and
Wna~, Wnb~,
227
n EN are increasing then (28)
if r < 1 then (rv28) holds. 0
Assume that r > 1 and let Cn ==
an bn b ' d n == b ' Un == W n ( an an+ n an+ n
+ bn) r ' Un
==
A
I.J.
un' n
*
E N .
Simple calculations then give:
snkl (g; YlJ + snkl (Q; 1Q) snkl (a + b; w)
=
snlr] (c· u) + sn[r] (d· u) < 2(.- 1 (c· u) + 2(.- 1 (d· u) n _,n _,- -
n
== 2ln
1
(
_,-
n
C+ d; U)
== 1.
by (26)
_,- '
0 COROLLARY
26 Ifr > 1 then for n > 1,
Wn ( (mkl (a; w)
+ mk1(b; w) r - (mk1(a+ b; w) r)
> Wn-1 ( (mt~ 1 (a; w) + mt~ 1 (b; w) 0
(29)
r - (mt~ 1 (a+ b; w) r).
Replace a, b by wtlr] (a; w), 9J1[r] (a; w) respectively in (28).
D
We now will generalize II 3.4 Lemma 11 and Corollary 13. LEMMA
27 If r, s E :IR, r < s then (30)
0
We can assume that r, s E JR* since the other cases follow from II 3.4 Lemma 11. Further we will assume that r == 1, s > 1, when (30) is
p(pas- (p- 1)bs) 1/s- (p- 1) (qbs- (q- 1)cs) 1/s s
> ( p (pa - (p - 1) b) - (p - 1) (qb - ( q - 1) c)
8
)1/s
.
(31)
228
Chapter III
Now if (p- l)r == (p- q)b + (q- 1)c the right-hand side of (31) can be written 8 )1/s s (p(pa(p1)b) (p1)(pb(p -1)!) , (
which by (28) is less than or equal to
So we will have proved (31) if we show that
equivalently
Is == (P- q b + q- 1 p-1 p-1
c) s 1 then
The proof follows that of II 3.4 Theorem 15, but using Lemma 27, rather
than II 3.4 Lemma 11. COROLLARY
D
29 If r, s E JR, r < s and n
> 1 then (32)
D
Immediate consequence of Theorem 28.
REMARK
(iii)
D
The above results are stated in [Nanjundiah 1952] with some extra
conditions on the weights, and was proved in the author's unpublished thesis that was communicated privately to the author who then published Nanjundiah's proof;
[Bullen 1996b]. A different proof given in [Rassias pp.27-37], [Mond & Pecaric 1996a,c; Tarnavas & Tarnavas] is based on the lemma of Kedlaya, VI 5 Lemma 5(a); [Kedlaya 1994, 1999]. The following deduction from (32) was made by Nanjundiah. COROLLARY
30
[HARDY's INEQUALITY]
9Jt~l (2t (b))
0, there is a unique s == so, so == rB, 0 < () < 1, at which ( Q~s (a; w)) rs attains its unique minimum value; if r < 0 there is a unique such s 0 at which the unique maximum is attained. THEOREM 2
E
D
Assume, without loss in generality, that the non-constant a is increasing, and put gr(s) == g(s) == log(Q~ 8 (a; w))rs. Simple calculations give:
If then r > 0, (C) gives that g" > 0 and so g1 is strictly increasing. Further
. g'( s ) == log ( Ln Wnarn r ) lim g'( s) == log ( LnWnaJ: r ) < 0; hm . W,;a. . W,;a. 2=1 " 'l ~=1 " 'l '( ) QJn(ar;w) ( ) g 0 == log ot ( ) < 0, by GA ; :«.n ar·, w _
s~-oo
s~oo
> 0·
( 1 ~ n Wiair 1og air r) g r == ""'n . r -log Wn ~·= Wiai > 0, .!.....Ji=l w'lai ., by (J), and the concavity of the logarithmic function. '( )
""'n .!.....Ji=I
1
'
Chapter III
232
These facts are sufficient to establish the result when r > 0; the case r
0 are: there is a unique minimum at
ro < s < s' then 9r(s) < 9r(s').
Since Q~ 8 (a; w)Q~,r(a; w) == 1 we can easily state a similar theorem
(iv)
with s fixed. REMARK
8
From the Remark (ii) if O
2
(Q~r(a;w))r ==
1, which is equivalent to (r;s). REMARK
< r < s then from ( Q~' 1 (a; w)) s > (Q~ 1 (a; w)) r, or Suppose that 1
(vi)
applied to 91:
the properties in Remark (ii),
a result that, in the equal weight case, can be found in [Berkolalko]. An alternative proof has been given in [Mitrovic 1970]. This is a particular case of the fundamental inequality between the Gini means, 5.2.1 (7). THEOREM
Assume that n
3
n-tuple with Wn
== 1, r, s
E
>
2, a an n-tuple with 0 < m < a < M, w an
JR, r < s, and let f3 == M/m, then
(2) where 8 _ (
r f3s _ 1 )
f3 S - j3T
1/r ({38
S
_ f3r
S -
T
)
j3T - 1
1/s
if rs-=/= 0,
'
8
1 ) 1 Is exp ( log f3 - -1 ) ' rr,s (!3) == ( r-+0slog f3 {3 8 - 1 s 1 lim rr,s (f3) == ( T log f3) /r exp (~ _ log j3 ) ' s-+0+ f3r - 1 r f3r - 1
ro,s (/3)
== lim
rr,O (f3)
==
f3
r
if r == 0, if s
== 0. (3)
Further if
B(r,s)
==
1 ( r s ) 8 s - r f3r - 1 f3 - 1 ' . 1 8(0, s) = hm B(r, s) = fJ r-+0s 1og
B(r, 0) = lim B(r, s) = s-+0+
1 r 1og
fJ
if rs-=/= 0,
1 j3S - 1' 1 j3r - 1'
if r
== 0,
if s
== 0,
(4)
233
Means and Their Inequalities
then 0 < B(r, s) < 1, and equality occurs in (2) if and only if for some set of indices I, Wr == (), ai == M,i E I and ai == m,i ¢:.I. D
Let us define
assuming, as will be shown, that this quantity does not depend on w. Writing A== {a; 1
< a < ,8}, W
Then using the basic properties of
rr,s (,B)
Wn == 1}, also define
== { w;
Q~ 8 ,
rr,s (,8; w)
== sup
< ,8,
w).
(5)
-aEA,wEW -
-wEW
Clearly rr,s (,8)
Q~s (a;
sup
==
and simple calculations establish the following identities:
(6) (rr,s (,B)) r == rl,s/r (,Br)' 0 < r < s < oo, (rr,s (,8)) - r
==
(ro,s (,B)) s == ro,l (,Bs)' 0 < s < oo, (7)
r-1,-s/r (,e-r), -()() < r < 0 < s
1;
(i)
(iii) r-l,t' 0 < t.
( ii) ro,l;
To evaluate rl,t (,B) let us first consider r 1,t (,8, w ).
Since the set A is compact there is a b E A such that
w))t (rl,t(,B ,_
==
(ffl)l,t(b· w))t ~n
-'-
==
I:~=l Wib~ . (""n . ·)t ui=l w~b~
(9)
n
For some k, 1 < k
< n,
put for s == 1 or t,
Wk ==
w,
L wibf == (1- w)a
8 •
Then if
i=l i=f.k
1
< x < ,8,
the function ¢,
has a maximum either at x == 1 or at x == ,8; this is immediate from a consideration of¢'. Since k, 1 < k
< n,
was arbitrary the point b of A given by (9) is such that
Chapter III
234
=
1, and if 0 < m < c1 1P' b-l/p < M
then
i=l
i=l
i=l
(21) with equality if and only if for all i, 1 < i The same inequality holds if p
D
< n, ci/p' b;l/p == m or M. < 0 but if 0 < p < 1 then (rv 21) holds.
This is a consequence of I 4.4 Theorem 23; see [PPT pp.114-115].
Let us now see how inequality (20) follows by taking equal weights and p (21). Put m 2 == m 2/M1, M 2 == M 2/m 1 in this special case of (21) to get
D
== 2 in
Chapter III
242 or
2
n
> 0.
Lcf-
>
i=1
This on rewriting gives (20).
A converse of (M) can be deduced from the converse of (H) given in Theorem 11; [Mond & Shisha 1967b].
14 Let b, c, m, M, /3, e, p be defined as in Theorem 11 but now with b1/P' (b + c) 1/P', c11P' (b + c) 11P' < M; then
THEOREM
m
1 then by Theorem 11 L(bi + ci)P i=1
n
n
n
==
L bi(bi + ci)P- 1 + L ci(bi + ci)P- 1 i=l
i=1
n
n
i=1
2=1
>E-1(Lbf)1/ p(L(bi + Ci)P)1/p' n
n
i=l
i=l
+ E-1 (L cf) 1/p (L(bi + Ci)P) 1/p'' which gives the result in this case. The other cases have similar proofs.
0
Means and Their Inequalities THEOREM
15 Let b and c ben-tuples, p > 1, 0
243
< m < b1 /p' c1 1P < M; then with
the notation of 4.2 Lemma 9,
with equality if and only if there is an index set, I, such that (1 - B) ~ b d bi1 Jp' ci-l/p == M , or m, accord.1ng as 2. E I , or 2. 'F d I L.iUf:.I iCi, an ,
LiEI
bici
==
e 0
This follows from 4.2 Theorem 10 in a manner similar to the way in which
Theorem 11 follows from 4.1 Theorem 3. REMARK
(v)
In the cases p
0
== 1, 2 converse inequalities of a different type have
been obtained by Benedetti who considers the maximum possible value when the n-tuples have terms that are restricted to certain finite sets of values. However it is beyond the scope of this work to consider the converses of (C), (H) and (M) in more detail. The reader is referred to the standard works [AI; BB; PPT] for further references in this area; see also [Dragomir & Gob 1997b; Izumino; Izumino
& Tominaga; T6th]. REMARK
(vi)
Converse inequalities have been based on the order relation of I 3.3;
see [Pecaric 1984a]. REMARK
(vii)
A converse of the equal weight case of 3.1.3 (9) has been given by
Toyama; see [AI p.285], [Toyama].
This was given a simpler proof and extended to the weighted case in [Alzer & Ruscheweyh 2000]. REMARK
(viii)
Leindler has proved that if 1 < p, q, r < oo, and if 1/p + 1/q
==
1 + 1/r then CX)
CX)
CX)
CX)
n=-CX)
n=-CX)
n=-CX)
m=-CX)
see [Leindler, 1972a,b,c, 1973a,b, 1976; Uhrin 1975]. The next result is a discrete form of a result of Zagier; [Alzer 1992e; Pecaric 1995b; Zagier].
Chapter III
244 THEOREM
16 Let a, b, c, d be sequences, a and b decreasing to zero, c, d
-
2 ~00 b2 Di=1 ai wi=1 i max{ A, B} ·
~00
> 1,
For any j
J
00
00
Laici == Caj i=l
+ L(ai- aj)ci < Caj + L(ai- aj)ci, i=1
~=1
so J
00
Dj Laici
< jCaj + D L(ai- aj) < max{C,D}Aj. i=1
i=l
Hence
00
LlbjDj L
aici
> max{C, D}AjLlbj.
which on summing over j gives the required result. REMARK
(ix)
D
Obviously from (C),
L
00
00
00
aibi < min {A, B, (L ai?1 2 (L bi?1 2 }. i=1
i=1
i=1
The following result is in [McLaughlin]. THEOREM
17 If a, b are real 2n-tuples then n
2n
2n
n
i=1
i=l
i=1
2 (La2ib2i-1-a2i-lb2i) < La;Lb;- (Laibi) . i=1
2
An inequality that is a mixture of (H) and (M) has been proved in [Iusem, Isnard & Butnariu]. THEOREM
18 If a and bare two n-tuples, p
> 2, define Ci ==
bf-
1
,
1
< i < n, then
Means and Their Inequalities If 1
245
< p < 2 the opposite inequality holds.
5 Other Means Defined Using Powers There are many other generalizations of the classical arithmetic, geometric and harmonic means besides the power means. Some generalizations are based on the very close connection of the power means with the convexity of certain functions; such generalizations are taken up in the next chapter. Other generalizations are really only defined in the case n
==
2 and these are considered in VI 2. Here we
study some generalizations that like the power means are based on the use powers, logarithms and exponentials.
5.1 COUNTER-HARMONIC MEANS DEFINITION 1 If a and w are n-tuples and r E JR. then the r-th counter-harmonic mean of a with weight w, is if r E JR., max a, . m1na,
if r if r
== oo, == -oo .
As with previous means we will just write iJ[rJ(a;w) when n reference is unambiguous,
(1)
== 2,
nkl
nk] (a) will denote the equal weight case, and if I
if the is an
index set the notation iJfJ (a; w) is used in the manner of I 4.2, II 3.2.2. The following identities are easily obtained: jj[l/ 2 ] (a, b)==
~[ ] (a· w) == 2(n (a· w)· ~[OJ (a· w) == -JJn ~ (a· w)· -'_ , - ' -JJn _,_,- '
REMARK
The reader can easily check that jj~] (a) is the point at which the
(i)
function I:~
2
1 ((ai- x)jx) takes its minimum value.
THEOREM 2 (a) If 1 < r
< oo then (2)
< r < 1 then (rv 2) holds. Inequality(2) is strict unless r == 1, oo or a
and if -oo is constant. (b) If -oo
< r < 0 then the following stronger result holds: (3)
246
Chapter III
Inequality (3) is strict unless r == -oo, 0 or a is constant. 0
The cases r
== ±oo are trivial so assume that r [r] (
)
E
JR; then
) ( 9J1k1 (g; 1Q) ) r-1 n a;w 9Jlk-ll(a;w)
9J1[r] (
,fjn a;w =
(4)
If 1 < r < oo then (r;s) and (4) imply (2), while if -oo < r < 1 (rv2) is implied. This completes the proof of (a). Now assume that r < 0 then:
nk1(a;w) == (flh--r+11(a-1;w))-1 < (9J1t-r+1](a-1;w))-1,
by(a),
== 9J1[r-1] (a. w). n _,_ D
This gives (b), and the cases of equality are immediate. REMARK
cases r
(ii)
Inequalities (2) and (3), together with 1 Theorem 2(c) justify the
== ±oo of Definition 1; see
REMARK
[HLP p. 62].
Inequality (2) in the equal weight case and with r
(iii)
== 2 is due to
Jacob; [Jacob]. THEOREM
3 If a and w are n-tuples and if -oo < r < s < oo then
(5) with equality if and only if a is constant . 0
(i) By 2.3 Lemma 6(b) and 2.3 Remark (iii), m(r)
==
2.:~ 1 wiar is strictly
log-convex if a is not constant; see also 3.1.1. So by I 4.1(3), if -oo < r < s < oo, log om(r) -log om(r- 1) 0, or
rk < 0; if s < r then F has a unique maximum at x == rk/(r- s), whereas the same point is the unique minimum if s > r. The methods of proof are those of elementary calculus; [Mitrinovic & Vasic 1966a]. The equal weight case with r == s had been considered earlier; see [Izumi, Kobayashi & Takahashi; Kobayashi]. Liu & Chen, [Liu & Chen], have considered the ratio
ryt(a w·r s·t) == _,_, ' '
(mkJ
(a; w)) t ----~ ( [r]( . ))t-1' 91tn a, w
which if s == t == r and, r - 1 is substituted for r is just ( 1), in the case of r finite. They then generalize (6) as follows. THEOREM 5
Ift
> 1,
and s
>1>r
9t(a + b, w; r, s; t) while if t
< 1,
and s
1 > q > 0 tben
#- q2
then inequality (7) is strict unless a, is constant.
(8) D
(a) The first limit is a simple use, after taking logarithms, of l'Hopital's
Rule 7 . As to the second: log ( lim
p-+~
QJ~'q(a; w))
== lim
1
p-+~p-q
(log(~ wiaf) -log(~ wian) ~ ~ ~=1
~=1
1 ~ == lim - log("' wiaf), p-+~p
== log(maxa),
~ ~=1
by 1 Theorem 2(c).
6 The contraharmonic mean is the solution x of the proportion x-a:b-x::b:a and is a nee-Pythagorean mean; see VI 2.1.4 7 See 1 Footnote 1.
Chapter III
250
The third limit is handled in a similar manner. (b) It is easily seen that if p =/= q
logoiB~'q(a;w)
= logom(p) -logom(q),
p-q where m is the function defined above in the proof of 5.1 Theorem 3, or in 3.1.1. This function is log-convex and so (7) in the case P1 =1- q1 and P2 =1- q2 is immediate from a basic property of convex functions, I 4.1 Remark (v), and the fact that the logarithmic function is strictly increasing; see [PPT pp.119-120]. The remaining cases follow by taking limits. (c) Assume that p
> q > 0 and write the right-hand side of (8) as,
_,_
n
_,_
(a· w))p/(p-q) ( 9Jt[P] n _,_
(9Jt[P] (b· n
w))pj(p-q)
_,_
(mt~l(a;w))qf(p-q) + (mt~l(b;w))qf(p-q).
Now apply Radon's inequality, 2.1 (9), in the case n == 2, and pin that reference taken as pj(p- q) to get Q)~,q (a;
Now by (M), p
W)
+ QJ~,q (b; W) >
+ 9Jt[P] (b· w))P/(p-q) n _,. (9J1~1(a; w) + 9J1~1(b; w))qj(p-q)
( 9Jt[P] (a· n
w)
_,-
> 1, and (rvM), 0 < q < 1, 9)1~] ({!; Yl) 9)1~1 (a; w)
+ 9Jtk1(fl.; Yl) > 91tW1(g +fl.; YL) + 9Jt~] (b; w)
- 9)1~1 (a+ b; w) ·
Hence
Q)p,q (a· w) n
The
othe~
REMARK
-'-
+ Q)p,q (b· w) > n -'- -
(9Jt[P] (a+
b· w) )P/(p-q) n _,== Q)p,q (a ( [ ] qj(p-q) n SJJt:f (a + b; w))
cases of (8) are easily proved.
(i)
+ -b·'w). 0
The proof of (b) is due to Pecaric & Beesack; [PPT p.119], [Pecaric
& Beesack 1986]. REMARK
(ii)
The case Pl
1, q1 == 0 of (7) has already been proved; see 4.1
Remark (v). REMARK
(iii)
Inequality (8) was first proved by Dresher and is sometimes referred
to as Dresher's inequality. The proof, which generalizes that of 5.1 (6), is due to Danskin; [PPT pp.120-121], [Danskin; Dresher]. A very detailed examination of
Means and Their Inequalities
251
== 2 has been made in [Losonczi & Pales 1996]. See also [B2
the case n
pp.233,
264], [Guljas, Pearce & Pecaric]. An extension of (8) is given in IV 7.1 Corollary 5. REMARK
(iv)
Note that Liapunov's inequality, 2.1 (8), is just
These means have been generalized to allow for complex conjugate p and q, [Pales 1989]. For another generalization see V 7.3. Gini means have been studied in detail by many authors; see for instance [Aczel REMARK
(v)
& Dar6czy; Allasia 1974-1975; Allasia & Sapelli; Brenner 1978; Clausing 1981;
Dar6czy & Losonczi; Dar6czy & Pales 1980; Farnsworth & Orr 1986; Jecklin 1948b; Jecklin & Eisenring; Losonczi 1971a,c 1977; Pales 1981, 1983a, 1988c; Persson & Sjostrand; Stolarsky 1996]. 5.2.2
Another kind of generalization has been suggested by
BoNFERRONI MEANS
Bonferroni, the Bonferroni means; [Bonferroni 1926, 1950]. Let a be ann-tuple and define
~p,q (a) == ( n
-
with obvious extensions to THEOREM
6
1 n(n- 1)
~~,q,r(a)
With the above notation if h > 0, q < p - h < p then ~~,q (a)
D 5.2.3
etc.
> ~~-h,q+h(a). D
For a proof the reader is referred to the references. GENERALISED PowER MEANS
In a very interesting paper Ku, Ku & Zhang
have considered a very general extension of power means ; [K u, K u & Zhang 1999]. Let ki, 1 < i < m, be distinct positive functions defined on the n-tuples a, and
== (k 1 , ... , km)· If w, such that I::n 1WiXi == 1. If r
write k
x are m-tuples, x allowed to be non-negative, and E lR the
r-th generalized power mean of a is :
1/r r) m , wixi(ki(a)) I:i=l (
(k.~ (a )) WiXi ' IT ~ ~=1
if r
-=1-
if r
== 0.
0,
The generalized power means include power means, certain Gini means, in particular the counter-harmonic means.
Chapter III
252 EXAMPLE
(i)
mk1 (a; w).
Take ki(a)
== ai and Xi== W;
1 ,
1
< i < n, when j{~n(a;k;w,x) == '
In fact, as we easily see,
(9) q EXAMPLE
(ii)
More generally take Xi =
we find that Jtk,nq](a;k;w,x)
L.::n ai
.
q,
i=l w~ai
== Q)k,q](a,w),
ki(a) = ai, 1 < i < n, when
p=j: q.
To obtain an extension of (r;s) here we need two further concepts. Firstly we define an order on tbe set of x; x
>>
x'
is a i 0 , 1 < io < m,
~there
such that
Xi < x~, io
+ 2 < i < m,
if i 0
+ 2 < m.
Secondly we define a condition of equality for tbe k: EQ(k, x
>> x', a)
holds if and
only if m
L
(ki(a)- ki 0 (a))
2
== 0 ~a is constant.
i=l
i=f.io ,xi =f.x~ 0
EXAMPLE
(iii)
then EQ(k, x
If for some distinct i, j we have ki (a)
>> x', a)
== kj (a)
holds for all distinct x, x', with x
~
a is constant
>> x'.
(iv) Particular cases of Example (iii) are given by: ki (a) == 9Jt[i] (a; w), d h _ m, an w ere _ 2. < vn (a ))n(ri-l)/ri(n-1) , £or l < :a-n (a )) (n-ri)/ri(n-1) (rlt and k ~·( a ) -_ (ot
EXAMPLE
0
< r1 < · · · < Tm < n.
THEOREM
7 (a) If r, s
E
R, r < s, tben
.ft~, n (a; k; w, X) < .ft~, n (a; k; w, X) , witb equality if and only if a is constant. (b) If k is decreasing, tbat is ki(a)
> ki+l (a),
1
< i < m, r > 0, and x
~
x' tben
(10) If x
0
# x'
and if EQ(k, x
~
x', a) is satisfied tben (10) is strict unless a is constant.
(a) This is immediate from (r;s) and (9).
253
Means and Their Inequalities (b) ( i) r == 0. Choose a {3 so that f3km
> 1.
Let i 0 be the suffix given in x
>>
x',
then since k is decreasing, 1
(f3ki)WiXi == (f3ki)WiX~(f3ki)Wi(Xi-X~) > (f3ki)WiX~(f3kio)Wi(;Ei-X~)
L::n
L::n
Further :L::n 1 WiXi == 1 WiX~ ==1 and so :L:~o 1 wi(xi-x~) == Using these facts we get that
io+ 1
< io.
wi(x~-xi)·
m
~o
II
{3Jt~],n(a; k;w,x) == II(f3ki)wixi(a)
(f3ki)wixi(a)
i=io+l
i=l
i=io+l
i=l
m i=io+l
m
II
>
io
(f3ki)wi(x~-xi)(a) II(f3ki)wix~(a) i=l
i=io+l
m
=={3
m
II
(f3ki)wixi(a)
i=io+l
w x'). II k "? x~ (a)- == {3Jt[o]m,n (a·-' -'k· -'1
~
i=l
Equality holds if fori=/= io, 1
x
~
< i < m,
x' there is some i such that Xi =/=
x~
(f3ki)wi(xi-x~)(a) == (f3ki 0 )wi(xi-x~). Since
and so the condition in the last sentence
implies that if i =/= i 0 then ki (a) == kio (a), and so by the condition EQ( k, x we must have that a is constant.
>> x', a),
( ii) r > 0 As the proof of this case is analogous to the one above it will be omitted. The D reader is referred to the reference for more details. REMARK
(i)
Particular cases of inequality (10) are inequalities 5.1(2),(5) and
inequality 5.2.1 (7). EXAMPLE ( v)
Let m == 3 and k1 (a) == 2tn (a), k2 (a) == Q)n (a), k3 (a) == SJn (a) and
take r =/= 0 in (10). If x
>> x'
then
11 (w1x12t~(a) + w2x2QJ~(a) + w3x 3 S)~(a)) r 11 > (w1x~2t~(a) + w2x~QJ~(a) + w 3 x~Sj~(a)) r.
The quantity here, J{r~[a; k; w, x], can be considered as a two parameter family of ' means lying between the extremes 2tn (a) and nn (a). Let a be ann-tuple, k and integer, 1 < k < n, and denote by aik), ... , a~) the ~, ~ == (~), k-tuples that can be formed from the elements of 5.3 MIXED MEANS
a.
Chapter III
254 DEFINITION
8 If s, t E 1R then the mixed mean of order s and t of a taken k at a
time is m1
ExAMPLE
(i)
n
(s ' t·' k·,_ a) ==
m1[s] (Wt[t] (a~k))
(i)
'
1 -< i < f£). -
(11)
The following special cases are immediate:
m1n(s, t; 1; a) == 9ltn(s, s; k; a)
REMARK
-~
k
K,
==
mk1 (a);
m1n(s, t; n; a)
== m-tkl (a).
An immediate consequence of (r;s) is that 9Itn(s, t; k; a) is an increas-
ing function of both s and t. The main results of this section are due to Carlson, Meany and Nelson; [Carlson 1970a,b; Carlson, Meany & Nelson 1971a,b]. THEOREM 9
If -oo
< s < t < oo
then
Wtn ( s, t; k - 1; a) < Wtn ( s, t; k; a) .
(12)
If s > t then ( rv 12) holds, and there is equality in both cases if and only if a is constant.
Denote by a~;), 1
< j < k, the collection of (k- 1)-tuples formed from the 1 elements of a~k). Then each a~k- ), 1 < h < f£ 1 == (k n 1), occurs (n- k + 1) times in the collection of a~;), 1 < j < k, 1 < i .::; ~"£; note that f£k == f£ 1 ( n - k + 1).
D
Firstly mt[t] (a~k)) == Wt(t] (m-t(t] (a~~)) 1 k -~ k k-1 -~J '
< · < k) ·' - J -
and so if s < t we have by (r;s) that mt[t] (a~) k
-~
> -
(a~.·) 1
9J1(s] (m(t]
-~J
k-1
k
'
< J. < k). -
(13)
Hence, by (11), and using the above notation,
!mn(s,t;k;a)
>9J1r1( !mr1 (9J1~~ 1 (afj),l <J < k), 1 < i < /\;) ==W1(s] (m(t] k
(ak-1) 1
k-1 -h
'
< h -< ~"£') -
==W1n(s, t; k- 1; a). If s
> t the inequality (13) is reversed. The cases of equality are immediate.
We now wish to obtain an inequality between different mixed means.
D
Means and Their Inequalities
255
Suppose that k +f > n; then a~k) and a]£) have m,m == m(i,j,k,f), elements in common, m =/= 0, 1
< i < j < A==
(~). For convenience we introduce the following
notations: a·
~
a·· ~J
== 9.J1[s] (a~k))· k
-~
'
1
== 9.J1(s] (a~k) n a\£))· k -~ -J '
LEMMA
If k
10
+f > n T · J
< rc,
1
< j 0 then:
¢(Wt~] (a; w)) + 2t3 (¢(a); w)
::> (w1 + w2)¢(Wt~] (a1, a2; w1, w2))
(3)
+ (w2 +w3)¢(Wt~1 (a2,a3;w2,w3)) + (w3 +w1)¢(9Jt~1 (a3,a1;w3,w1)). The methods of proof were elementary and the paper caused much interest. In particular if in the second case r
== 1, when the mean is the arithmetic mean, the
condition on ¢ implies that it is convex. This simpler condition has been used to obtain the same result, see [Baston 1976; Vasic & Stankovic 1976], and [Pecaric 1986] for a different proof. It was Pecaric who noted that (3), with r
the following more general result; [PPT pp.111-180].
== 1, implies
Chapter III
260 THEOREM
a
< ai < b,
5 If¢ : [a, b] ~--+ lR is convex, w an n-tuple, a a real n-tuple with 1
< i < n,
and if 2
m. Then lim
an==
m
(
m2+1)
) 1 Iron-[ r] ( ·1 I r ;.~"'-m 'l
ai'
1
(1Bm(aLloo
D
(
< .< -
)
m '
'l -
1
'f 1
r
_j_
I
0
'
ifr=O.
),
(a) See [Math. Lapok, Problem F 1930, 50 (1975), 11-12].
The case p
>
==
2 is in [Mitrinovic, Newman & Lehmer]; see also [AI p.340], 1 was stated without proof in [Ozeki 1968]; a proof was given in
(b) The case n
[Mitrinovic & Kalajdzic]. A proof for p < 1 is given in [Russell]. (c) This is a generalization, by Vasic, of a result of Markovic. A different proof has been given by Pecaric who also shows that the same inequality holds if f(x 1 1k) is convex on [0, ak]; [Keckic & Lackovic; Markovic; Pecaric 1980a; Vasic 1968]. (d) See [Elem. Math. 29 (1974), 15-16].
> 2 and w is an n-tuple, Wn m < a < M, and if r < s, t =/= 0, then THEOREM 8
1 M(s-r)
If n
1, a a real n-tuple with
r(r- t)) < (9Jt~l(g;IQ)r- (9Jt~l(g;IQ)r -
(5)
'
with equality if and only if a is constant. D
This follows from II 3.8 Theorem 27 by taking x
==
0 and x
==
r.
D
II 3. 7 Corollary 28 is the case t == 1 of this theorem. As in that case we note that it is sufficient to assume that the n-tuples a, w be similarly ordered, (ii)
REMARK
see II 3.8 Remark (viii). (iii)
Further interesting properties of the left-hand side of (5) can be found in [Pecaric 1988; Pecaric & Ra§a 1993]. REMARK
In the case of the special sequence v == {1, 2, ... } certain special results are obtainable of which the most interesting is perhaps the following. THEOREM
10 Ifr > 0 and n > 1 then
9J1[r] (v)
n
--
1. Further we can, without loss in generality,
assurne that a is increasing. Let V = M[t] (M, rn; a) - M[l] (M, rn; a) then simple calculations show that:
8V/ 8a;
= Wi ( 1- (a;jM[tl) t-l), and BIM[tJ_ail/ 8ai = ±( wia~-l (M[t] )(1-t)/t+ 1),
according as a.i >
M[t], a.i
0, ry
#
1, taking M(x) == 1 1 /x gives another family of means,
called the radical means:
ExAMPLE
(iii)
If M(x) == xx, x >
1 e- ,
we get the so-called basis-exponential
mean; if its value is 11 then
ExAMPLE
·(iv)
If M(x) == x 11x, x
>
1 e- ,
then in a similar manner we get the
basis-radical mean; if its value is v then n
v
1/v _ _1_
-
w;
"'"""' ~
n.
. 1/ ai w 2 ai .
1.=1
REMARK
(ix)
The above means are some of the many introduced and used by
the Italian school of statisticians; see [Bonferroni 1923-4,1924-5,1927; Gini 1926; Pizzetti 1939; Ricci].
The following results can be obtained.
270
Chapter IV
3
THEOREM
If a is a positive n-tuple then
n n n n [ J n 1)!2tn(a)ti < l:af